SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks
Extracted source text · arXiv 2606.27807
%%%%%%%% ICML 2026 EXAMPLE LATEX SUBMISSION FILE %%%%%%%%%%%%%%%%%
\documentclass{article}
% Recommended, but optional, packages for figures and better typesetting:
% \usepackage{amsmath}
\usepackage{microtype}
\usepackage{graphicx}
\usepackage{subcaption}
\usepackage{tabularx}
\usepackage{booktabs}
\usepackage{multirow}% for professional tables
\usepackage{pifont}
% hyperref makes hyperlinks in the resulting PDF.
% If your build breaks (sometimes temporarily if a hyperlink spans a page)
% please comment out the following usepackage line and replace
% \usepackage{icml2026} with \usepackage[nohyperref]{icml2026} above.
\usepackage{hyperref}
% Attempt to make hyperref and algorithmic work together better:
% \newcommand{\theHalgorithm}{\arabic{algorithm}}
% Use the following line for the initial blind version submitted for review:
% \usepackage{icml2026}
\usepackage[accepted]{icml2026}
% For preprint, use
% \usepackage[preprint]{icml2026}
% If accepted, instead use the following line for the camera-ready submission:
% \usepackage[accepted]{icml2026}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{mathtools}
\usepackage{amsthm}
% if you use cleveref..
\usepackage[capitalize,noabbrev]{cleveref}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% THEOREMS
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\theoremstyle{plain}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{assumption}[theorem]{Assumption}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
% Todonotes is useful during development; simply uncomment the next line
% and comment out the line below the next line to turn off comments
%\usepackage[disable,textsize=tiny]{todonotes}
\usepackage[textsize=tiny]{todonotes}
% The \icmltitle you define below is probably too long as a header.
% Therefore, a short form for the running title is supplied here:
% \icmltitlerunning{Submission and Formatting Instructions for ICML 2026}
\icmltitlerunning{SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks}
\begin{document}
\twocolumn[
\icmltitle{SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks}
% It is OKAY to include author information, even for blind submissions: the
% style file will automatically remove it for you unless you've provided
% the [accepted] option to the icml2026 package.
% List of affiliations: The first argument should be a (short) identifier you
% will use later to specify author affiliations Academic affiliations
% should list Department, University, City, Region, Country Industry
% affiliations should list Company, City, Region, Country
% You can specify symbols, otherwise they are numbered in order. Ideally, you
% should not use this facility. Affiliations will be numbered in order of
% appearance and this is the preferred way.
\icmlsetsymbol{equal}{*}
\begin{icmlauthorlist}
\icmlauthor{Ruiqi Song}{equal,tj,casia,space}
\icmlauthor{Dujun Nie}{equal,casia,space,cas}
\icmlauthor{Siyu Teng}{space,sz}
\icmlauthor{Baiyong Ding}{casia,waytous}
\icmlauthor{Xiaotong Zhang}{casia,space,cas}
\icmlauthor{Dong Li}{space,Macau}
\icmlauthor{Chenming Zhang}{space,waytous,IAIR}
%\icmlauthor{}{sch}
\icmlauthor{Yuchen Li}{space,Munich}
\icmlauthor{Hangbin Wu}{tj}
\icmlauthor{Long Chen}{casia,space,waytous,IAIR}
%\icmlauthor{}{sch}
%\icmlauthor{}{sch}
\end{icmlauthorlist}
\icmlaffiliation{tj}{College of Surveying and Geo-informatics, Tongji University, Shanghai, China}
\icmlaffiliation{casia}{State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China}
\icmlaffiliation{IAIR}{IAIR, Xi'an Jiaotong University, Xi’an, China}
\icmlaffiliation{waytous}{Waytous, Wuhan, China}
\icmlaffiliation{Munich}{Department of Informatics, Technical University of Munich, Munich, Germany}
% \icmlaffiliation{Obuda}{University Research and Innovation Center, Obuda University, Budapest, Hungary}
\icmlaffiliation{sz}{Shenzhen University, Shenzhen, China}
\icmlaffiliation{cas}{School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China}
\icmlaffiliation{Macau}{Faculty of Innovation Engineering, Macau University of Science and Technology, Macau, China}
\icmlaffiliation{space}{OpenSpace Lab, China}
\icmlcorrespondingauthor{Hangbin Wu}{hb@tongji.edu.cn}
\icmlcorrespondingauthor{Long Chen}{long.chen@ia.ac.cn}
% You may provide any keywords that you find helpful for describing your
% paper; these are used to populate the "keywords" metadata in the PDF but
% will not be shown in the document
\icmlkeywords{Machine Learning, ICML}
\vskip 0.3in
]
% this must go after the closing bracket ] following \twocolumn[ ...
% This command actually creates the footnote in the first column listing the
% affiliations and the copyright notice. The command takes one argument, which
% is text to display at the start of the footnote. The \icmlEqualContribution
% command is standard text for equal contribution. Remove it (just {}) if you
% do not need this facility.
% Use ONE of the following lines. DO NOT remove the command.
% If you have no special notice, KEEP empty braces:
\printAffiliationsAndNotice{\icmlEqualContribution}
% \printAffiliationsAndNotice{} % no special notice (required even if empty)
% Or, if applicable, use the standard equal contribution text:
% \printAffiliationsAndNotice{\icmlEqualContribution}
\input{section/00_abstract}
\input{section/01_intro}
\input{section/02_related}
\input{section/03_method}
\input{section/04_experiment}
\input{section/05_conclusion}
\section*{Acknowledgements}
This work was supported by the Joint Funds of the National Natural Science Foundation of China under Grant U24B20162 and the New Generation Artificial Intelligence-National Science and Technology Major Project (No. 2025ZD0124205).
\section*{Impact Statement}
SpikeVLA integrates spiking neural networks with vision-language-action models, enabling energy-efficient, brain-inspired multi-modal reasoning and control. This approach can support safer and more robust autonomous agents, assistive robotics, and sustainable AI applications. Careful attention to fairness and bias is essential to ensure responsible deployment in human-facing, high-stakes contexts such as healthcare and education.
\nocite{langley00}
\bibliography{bib/11_references}
\bibliographystyle{icml2026}
\input{section/appendix}
% \input{rebuttal}
\end{document}
\begin{abstract}
% Vision-Language-Action (VLA) 模型作为具身智能的重要范式，能够将视觉感知、语言理解与动作决策有机结合。然而，现有 VLA 模型普遍依赖大规模 Transformer 结构，计算开销与能耗过高，限制了其在低功耗与实时场景中的应用。为此，我们提出 SpikeVLA，首个基于 Spiking Neural Networks (SNNs) 的 VLA 框架，以实现轻量化与节能化的具身智能。SpikeVLA 具有三项核心创新：(i) Spiking Vision-Language Reasoning (SVLR) ——提出脉冲驱动的多模态激活机制，以事件驱动方式选择性稀疏激活视觉与语言通道，大幅降低大模型推理能耗；(ii) Spiking Neural Network for Action Policy ——设计基于群体编码的脉冲化策略网络，并结合强化学习进行优化，在连续控制任务中实现更高能效与鲁棒性；(iii) Energy-Efficient Benchmark for VLA ——提出首个面向 VLA 的能效评测协议与指标体系，覆盖长指令–短路径、短指令–长路径和密集障碍物导航等压力测试场景，并结合标准化指标（精度、延迟、能耗与能效比），为不同架构提供可复现的系统性比较。在多模态交互与机器人控制任务上的实验结果表明，SpikeVLA 在保持与传统 ANN 基线相当性能的同时显著降低了能耗，展现了在低功耗、实时具身智能中的应用潜力。本研究为构建可持续的 neuromorphic VLA 系统提供了新的范式。
Vision-Language-Action (VLA) models have become a dominant paradigm for embodied intelligence. However, most existing approaches are built on large-scale transformers, resulting in substantial inference latency and energy consumption that limit their practical deployment in low-power, real-time scenarios. We propose SpikeVLA, a spiking VLA architecture for embodied navigation with energy-efficient inference, consisting of three key components. (i) A spiking vision encoder, Spike-V, that replaces dense continuous layers with event-driven spiking layers to reduce the energy consumption of visual representation learning. (ii) A multi-modal spiking large language model, Spike-L, that reformulates cross-modal reasoning with spiking dynamics and token-level event-driven sparsity to further lower computational cost. (iii) A spiking action policy network, Spike-A employs Laplacian-kernel population coding with a multi-layer fully connected SNN, and decodes spiking activities into stable and robust continuous control with energy-efficient inference under low-power constraints.
Experiments on navigation and robotic control tasks show that SpikeVLA significantly reduces energy consumption and computational cost while maintaining competitive performance, highlighting its potential for low-power, real-time embodied intelligence.
% \vspace{-5mm}
\end{abstract}
\section{Introduction}
\label{sec:intro}
% \includegraphics[width=0.9\textwidth]{section/figure/SpikeVLA_02.png}
% % \vspace{-2mm}
% \captionof{figure}{SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks.
% % We also release SceneVerse-597k-R1, a multi-agent interaction and reasoning dataset.
% }\label{fig:overview}
% To insert a figure: \input{figs/template}
% Or table: \input{tables/template}
% 第一段：背景与重要性
% 近年来，Vision-Language-Action (VLA) 模型逐渐成为具身智能研究的重要方向。通过统一视觉感知、语言理解与动作决策，VLA 模型能够在复杂环境中实现指令驱动的自主导航与任务执行。这类方法在机器人控制、交互式导航和多模态理解等任务中展现出优异的性能，为通用型具身智能系统奠定了基础。然而，随着模型规模不断增长，VLA 在推理开销与能耗方面的不足愈发明显，严重制约了其在移动机器人、无人机和腿足机器人等低功耗平台上的落地应用。因此，探索 高能效、轻量化的 VLA 模型 已成为具身智能发展的关键需求。
% 第二段：现有方法的局限性
% 现有代表性工作（如 NaVILA）多采用大规模 Transformer 结构结合 ANN 策略网络来完成多模态推理与动作决策。虽然这类架构在任务成功率和泛化能力上表现良好，但其 密集计算模式 导致能耗过高，且难以在实时性要求严格的场景中保持稳定性能。更为重要的是，当前的 VLA 基准主要关注任务完成率、路径效率等指标，而 缺乏对能耗和延迟的系统性评估，使得现有方法难以在低功耗、实时控制任务中获得可靠的对比。基于此，如何同时保证 任务性能 与 能效优势，成为亟待解决的核心挑战。
% 第三段：我们的思路
% 为应对上述问题，我们提出 SpikeVLA，这是首个基于 Spiking Neural Networks (SNNs) 的端到端 VLA 框架。SNN 通过事件驱动的稀疏激活机制天然具备高能效特性，非常适合资源受限和低功耗场景。在 NaVILA 的基线上，我们在两个关键环节引入了脉冲化设计：一方面，在感知–语言推理端提出 Spiking Vision-Language Reasoning (SVLR)，利用脉冲激活机制选择性稀疏化视觉与语言 token，从而大幅降低大模型推理能耗；另一方面，在动作控制端提出 Spiking Neural Network for Action Policy，通过群体编码与强化学习训练脉冲化策略网络，使其在连续控制任务中兼具能效与鲁棒性。由此，SpikeVLA 实现了从感知到控制的端到端能效优化。
% 第四段：能效评测基准的提出
% 仅有方法创新不足以全面揭示 SNN 在 VLA 中的优势。现有 VLA 评测体系忽视了能耗与实时性这一关键维度，导致研究成果缺乏系统性比较。为此，我们进一步提出 Energy-Stress Benchmark for VLA，构建三类能效敏感场景：长指令+短路径（推理负载瓶颈）、短指令+长路径（动作执行瓶颈）、以及 密集障碍物导航（端到端实时性瓶颈）。结合标准化协议与统一指标（如 J/step、J/episode、Energy-Delay Product），该基准能够在不同任务压力下全面刻画模型的能效表现，从而为 ANN 与 SNN 架构提供公平可复现的对比。
% 第五段：总结与贡献
% 综合来看，本文的贡献主要体现在三个方面：
% 我们提出 SpikeVLA，首个基于 SNN 的端到端 VLA 框架，面向低功耗与实时具身智能场景。
% 我们设计了两项关键模块：Spiking Vision-Language Reasoning 和 Spiking Neural Network Policy，分别在感知–推理与动作控制端实现能效化改进。
% 我们建立首个 能效压力测试基准，覆盖不同能耗瓶颈场景，并提供标准化指标体系，实现对 VLA 模型的系统性能效评估。
% 实验结果表明，SpikeVLA 在保持与 ANN 基线相当性能的同时显著降低了能耗，展示了 neuromorphic computing 在 VLA 模型中的应用潜力。
\begin{figure}[t]
\centering \includegraphics[width=1\columnwidth]{section/figure/overview_c.pdf}
% \vspace{-0.5em}
\caption{SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks.}
\label{fig:overview}
% \vspace{-24pt}
\end{figure}
In recent years, Vision-Language-Action (VLA) models have become increasingly prominent as a key paradigm for embodied intelligence, unifying vision, language, and action to support navigation, manipulation, and collaboration tasks in complex environments. Despite significant progress in embodied intelligence, the rising inference-time compute footprint and energy demands of large VLA models increasingly constrain on-board deployment on resource-limited platforms (e.g., micro-robots, legged systems, and deep space exploration robots) and latency-critical real-time operation, which is particularly challenging for space exploration robotics. Therefore, lightweight and energy-efficient VLA models are critical for the scalable real-world deployment of embodied intelligence.
Recent VLA models such as NaVid~\cite{navid}, NaVILA~\cite{navila}, and UniNaVid~\cite{Uni-navid} typically rely on large-scale transformers for multimodal reasoning, together with ANN-based policy networks for control. Although they perform strongly on instruction-driven navigation, their dense computation imposes considerable energy and latency costs, limiting deployment on resource-constrained platforms and in latency-critical scenarios.
To address this challenge, previous work has explored efficiency-oriented designs. For example, COSMO~\cite{COSMO} reduces inference cost through selective memorization, while VL-Nav~\cite{VL-Nav} focuses on efficient spatial reasoning with real-time deployment.
Nevertheless, these methods are still grounded in continuous and dense computation, which offers limited headroom for further efficiency improvements. In contrast, event-driven spiking neural networks introduce a fundamentally different, energy-aware paradigm for computation. They adopt an event-driven sparse spiking mechanism that triggers computation only when information changes, significantly reducing redundant operations and energy consumption, which introduces a new paradigm for low-power VLA deployment.
To address these challenges, we propose SpikeVLA, the first VLA architecture built on spiking neural networks, which represents a trade-off between performance and efficiency, as shown in Fig.~\ref{fig:overview}.
% Unlike ANNs, SNNs achieve energy-efficient computation through event-driven sparse activations.
SpikeVLA consists of three complementary modules. Spike-V provides energy-efficient visual representations through event-driven spiking visual encoding. Spike-L reformulates multimodal reasoning with spiking dynamics and token-level sparsity to reduce inference-time compute and energy cost. Spike-A achieves stable and robust continuous control with a fully spiking action policy, supporting low-energy operation. Together, these components establish SpikeVLA as a neuromorphic computing paradigm for embodied intelligence.
In summary, this work makes three main contributions.
\textbf{1)} We propose SpikeVLA, the first VLA framework built on spiking neural networks. Through event-driven sparse activations, SpikeVLA substantially reduces energy consumption in reasoning and control, shifting VLA from dense continuous computation toward a low-power, real-time neuromorphic paradigm.
\textbf{2)} We develop a three-module spiking architecture: Spike-V for event-driven spiking visual encoding, Spike-L for energy-efficient multimodal reasoning with token-level sparsity, and Spike-A for stable continuous control with a fully spiking policy network.
\textbf{3)} We conduct extensive experiments on Vision-Language Navigation and robotic control tasks, demonstrating that SpikeVLA matches the performance of ANN-based baselines while reducing energy consumption and computational cost, and serves as a general energy-efficient design paradigm transferable to other robotic VLA tasks.
\noindent\textbf{Conflict of Interest Disclosure.}
The authors declare that they have no financial conflicts of interest related to this work. All research and writing were conducted independently and solely for academic purposes.
% \begin{figure}[t]
% \centering \includegraphics[width=\columnwidth]{section/figure/ANNvsSNN2.pdf}\vspace{-0.5em}
% \caption{SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks.}
% \label{fig:overview}
% \vspace{-24pt}
% \end{figure}
\section{Related Work}
\label{sec:related}
\subsection{Vision-Language-Action for Navigation}
Early vision-and-language navigation methods in continuous environments relied on end-to-end action prediction, as represented by Seq2Seq~\cite{Seq2Seq}. However, these methods often suffer from error accumulation and semantic drift in long-horizon rollouts.
Some approaches use waypoint-based and hierarchical planning to decouple high-level intent from low-level execution~\cite{HPN+DN, law, AG-CMTP}.
% Some other works explicitly model route structure through topological planning, as in AG-CMTP~\cite{AG-CMTP}, shifting from reactive policies toward plan-and-act pipelines.
Moreover, subsequent research has increasingly focused on the generalization of the cross-scene and the integration of spatial priors~\cite{CMA, VLNBERT, sim2sim}. In parallel, methods such as CM2~\cite{cm2} and WS-MGMap~\cite{WS-MGMap} have been proposed. For longer horizons and more complex environments, recent approaches have increasingly focused on global state maintenance, memory, and online closed-loop operation~\cite{gridmm, ego-map, dreamwalker, hamt}. Recent research has increasingly focused on enhanced planning, representations, unified modeling, and streaming closed-loop navigation~\cite{etpnav, hnr, navid, Uni-navid, instructnav, streamvln, CorrectNav}. However, this trend also increases the computational cost of continuous inference during long-horizon closed-loop operations.
As a result, the demand for low-power, real-time VLA deployment is increasing, driving the need for efficiency-oriented designs~\cite{COSMO, VL-Nav}. However, these methods primarily operate within a continuous, dense-computation paradigm, which limits further efficiency gains and highlights the need for neuromorphic, event-driven VLA models like SpikeVLA.
% Along this trajectory, SpikeVLA targets the energy and latency constraints of streaming long-horizon deployment by introducing event-driven spiking modeling to reduce continual inference overhead, complementing ongoing trends in planner enhancement and self-correction.
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{section/figure/spike_framework.pdf}
% \vspace{-2em}
\caption{\textbf{Architecture of SpikeVLA.} We introduce an SNN-based VLA architecture composed of a spiking neural network vision encoder, a multimodal spiking large language model, and an SNN action policy. Under a unified event-driven paradigm, this design establishes a new architectural paradigm for VLA.}
\label{fig:spike_framework}
% \vspace{-18pt}
\end{figure*}
\subsection{Spiking Neural Networks}
Spiking neural networks (SNNs) are promising for efficient temporal learning, thanks to their event-driven computation and potential for energy savings.
Many works use ANN-to-SNN conversion to preserve ANN capabilities while facilitating low-power inference, including~\cite{sign, inference, towards, differential}.
In parallel, efficient spiking architectures have been developed to improve vision performance and provide general building blocks for multimodal and long-horizon scenarios~\cite{spike2former,staa}.
Building on these foundations, recent works have extended spiking models to foundation models through generative pretraining, distillation, enhanced spiking mechanisms, and more efficient spike allocation~\cite{spikingbert, spikelm, spikellm}.
% Additionally, spiking pretraining has been applied to vision-language representation learning~\cite{svl}.
More recently, SNNs have been increasingly explored in embodied domains such as continuous control. Previous works have enhanced the expressivity of spiking policies through fully spiking actor designs, while also addressing the mismatch between discrete spiking dynamics, such as Proxy Target~\cite{exploring, multi, proxy}.
Applications in autonomous driving further demonstrate the feasibility of SNNs for long-horizon tasks under the efficiency constraints platform~\cite{autonomous}.
Inspired by these developments, SpikeVLA is designed for energy-constrained deployment, using an event-driven spiking architecture to reduce inference costs.
% \begin{figure}[t]
% \centering
% \includegraphics[width=1\textwidth]{section/figure/ANNvsSNN2.pdf}
% \vspace{0em} \caption{\textbf{Ablation of Critic Network Dimensions.} }
% \label{fig:ann2snn}
% % \vspace{16pt}
% \end{figure}
\section{Method}
\label{sec:method}
% \subsection{Problem Formulation}
\subsection{Architecture}
We propose SpikeVLA, an end-to-end spiking VLA architecture for embodied navigation, consisting of a spiking vision encoder (Spike-V), a multimodal spiking language model (Spike-L), and a fully spiking action policy network (Spike-A), as illustrated in Fig.~\ref{fig:spike_framework}. Spike-V takes the current frame along with a sequence of past frames, encoding temporal observations into event-driven spiking visual tokens for energy-efficient visual encoding (see Section~\ref{sec:Spike-V}). Spike-L fuses visual and text tokens and performs event-driven channel-wise sparsification to reduce computational cost (see Section~\ref{sec:Spike-L}). Spike-A maps the fused multimodal representation to continuous actions with a fully spiking policy network, achieving stable and robust closed-loop control under low-power constraints (see Section~\ref{sec:Spike-A}).
\subsection{Spike Neural Network Vision Encoder}
\label{sec:Spike-V}
We propose an SNN-based visual encoder that fuses the current frame with history frames to provide temporal context for time-dependent VLA tasks. Built upon a spiking Transformer, it replaces dense continuous feature extraction with sparse, spike-driven computation over discrete timesteps. We reformulate linear layers and key nonlinear operations in spiking form, reducing computation and energy while preserving representational capacity for efficient temporal visual feature extraction.
\noindent\textbf{Differential Spiking Neurons.}
To preserve high-fidelity representations with few timesteps $T$, we adopt differential coding to express continuous activations as incrementally accumulated updates over time, so that activation updates in both linear mappings and nonlinear transformations follow a unified incremental recursion. This process can be written as follows:
% \begin{equation}
% \label{eq:diff_coding}
% \begin{gathered}
% e^{l}[t] = r^{l}[t-1] + x^{l}[t],\\
% r^{l}[t] = r^{l}[t-1] + \frac{x^{l}[t]}{t}, \quad
% r^{l}[0] = 0.
% \end{gathered}
% \end{equation}
\begin{equation}
% \tag{13}
\begin{gathered}
\delta^{l}[t] \;=\; \bar{a}^{l}[t-1] \;+\; \theta^{l}\, z^{l}[t] \\
\bar{a}^{l}[t] \;=\; \bar{a}^{l}[t-1] \;+\; \frac{\theta^{l}}{t}\, z^{l}[t]
\;=\;\frac{1}{t}\sum_{i=1}^{t}\delta^{l}[i],
% \vspace{-3mm}
\end{gathered}
\end{equation}
where $l$ denotes the layer index and $\theta^{l}$ is the firing threshold of spiking neurons in layer $l$. $z^{l}[t]\in\{0,1\}$ indicates whether a spike is emitted at time step $t$.
$\delta^{l}[t]$ represents the encoded output produced by differential coding, while $\bar{a}^{l}[t]$ is the average of $\delta^{l}[t]$ from time $1$ to time-step $t$ with initialization $\bar{a}^{l}[0]=0$.
% Here, $x^{l}[t]$ denotes the incremental correction signal at time step $t$, $r^{l}[t]$ is the accumulated coded state up to $t$, $e^{l}[t]$ is the instantaneous output decoded from the previous state and the current increment, and the $1/t$ scaling factor stabilizes updates at early time steps.
Building on differential coding, we instantiate spiking units as differential spiking neurons, supporting stable spiking dynamics and sparse event-driven transmission with only a few timesteps $T$. We introduce a recurrent auxiliary state into membrane updates to incrementally correct the input current based on both input changes and emitted spikes, aligning neuron dynamics with the differential-coding recursion and promoting sparse event-driven transmission. The mechanism is formulated as follows:
% \begin{equation}
% \label{eq:diff_spiking_neuron_theta}
% \begin{aligned}
% I^{l}[t] &= m_r^{l}[t] + x^{l-1}[t],\\
% m_r^{l}[t+1] &= m_r^{l}[t] + \frac{x^{l-1}[t]}{t} - \frac{\theta^{l} z^{l}[t]}{t},
% % \qquad z^{l}[t]\in\{0,1\}.
% \end{aligned}
% \end{equation}
\begin{equation}
\label{eq:diff_spiking_neuron_theta}
\begin{aligned}
I^{l}[t] &= c_r^{l}[t] + x^{l-1}[t],\\
c_r^{l}[t+1] &= c_r^{l}[t] + \frac{x^{l-1}[t]}{t} - \frac{\theta^{l} z^{l}[t]}{t},
\end{aligned}
% \vspace{-2mm}
\end{equation}
where $I^{l}[t]$ is the input current of the differential spiking neuron at time step $t$, and $x^{l-1}[t]$ is the output of the previous layer. $c_r^{l}[t]$ denotes the recurrent residual state for the input-current correction. $\theta^{l}$ is the firing threshold of the layer $l$, and $z^{l}[t]\in\{0,1\}$ indicates whether a spike is emitted at the time step $t$, producing the output $x^{l}[t]=\theta^{l}z^{l}[t]$.
Fully connected, convolutional, and matrix multiplication layers constitute the main computational bottlenecks, and inserting spiking neuron layers before them supports event-driven processing and reduces energy consumption.
% Here, $x[t]$ and $y[t]$ denote the synaptic input and spike output at time step $t$, respectively, and $I[t]$ is the effective input current for membrane potential updates. The auxiliary state $m_r[t]$ accumulates a $1/t$-scaled correction from both the synaptic input and the emitted spikes, enforcing consistency with the differential coding recursion and stabilizing early-step updates.
% This suppresses redundant firing and promotes sparse event-driven spike transmission.
\noindent\textbf{Linear-Layer Conversion.}
We model compute-intensive mapping operations as linear layers and drive their computation with time-step increments generated by differential spiking neurons, resulting in event-driven execution.
% A general linear transformation can be expressed as:
% \begin{equation}
% \label{eq:linear_map}
% x^{l} = W^{l} x^{l-1} + b^{l},
% \end{equation}
% where \(W^{l}\) is the weight matrix, \(x^{l-1}\) is the input from the previous layer, \(b^{l}\) is the bias, and \(x^{l}\) is the output at layer \(l\).
Each neuron applies a linear mapping to its input at every time step. Equation~\eqref{eq:linear_step} specifies the membrane-potential update for neurons in the $l$-th layer. To prevent bias accumulation, the bias term is moved into the initial membrane potential of the next unit. It is expressed as follows:
\vspace{-2pt}
\begin{equation}
% \vspace{-2mm}
\label{eq:linear_step}
x^{l}[t] = W^{l} x^{l-1}[t],
% \vspace{-1mm}
\end{equation}
Where \(x^{l}[t]\) is the output of the \(l\)-th layer at time step \(t\), \(W^{l}\) is the weight matrix for the \(l\)-th layer, and \(x^{l-1}[t]\) is the input of the \((l-1)\)-th layer at time step \(t\).
We build our visual encoder on SigLIPv2 and apply the proposed operation to linear layers, such as MLP blocks and attention projection matrices. This converts matrix multiplications and convolutions from dense continuous computation to event-driven operations.
\noindent\textbf{Nonlinear-Layer Conversion.}
Beyond linear mappings, Transformers further require nonlinear operators such as normalization, activations, and attention computations that combine normalization with multiplicative interactions. To integrate these operators into time-step inference while preserving their functional forms, we employ differential graded units. The key idea is to output the incremental change at each time step, avoiding redundant recomputation and transmission of the full continuous value. This enhances fidelity under small $T$ conditions while not requiring additional fitting or retraining.
For any single-input nonlinear operator \( F(\cdot) \), we realize its dynamics with a differential graded unit under differential coding during the ANN-to-SNN conversion. This mapping is given by the following equations, which convert continuous nonlinear operators into graded updates to improve efficiency and avoid redundant computation.
% \vspace{-6pt}
\begin{equation}
% \vspace{-3mm}
\label{eq:graded_single}
\begin{gathered}
c^l[t] = c^l[t-1] + \frac{x^{l-1}[t]}{t},\\
\Delta_F^l[t] = F^l(c^l[t]) - F^l(c^l[t-1]),\quad
x^l[t] = t\,*\Delta_F[t].
\end{gathered}
% \vspace{-1mm}
\end{equation}
where $F^{l}(\cdot)$ denotes the nonlinear operator in layer $l$ that takes a single input $x^{l-1}[t]$, and $c^{l}[t]$ denotes the membrane potential at time step $t$.
During ANN-to-SNN conversion, differential coding provides a dynamic realization for operators with two inputs. We maintain two states, \( c_a[t] \) and \( c_b[t] \), and define the mapping as follows:
\vspace{-5pt}
\begin{equation}
% \vspace{-1mm}
\label{eq:graded_binary}
\begin{gathered}
c^l_a[t] = c^l_a[t-1] + \frac{x^{l-1}_a[t]}{t},\\
c^l_b[t] = c^l_b[t-1] + \frac{x^{l-1}_b[t]}{t}, \\
% \end{gathered}
% \vspace{-1mm}
% \end{equation}
% \begin{equation}
% \begin{gathered}
x^l[t] = t*\Big(c^l_a[t]\odot c^l_b[t]-c^l_a[t-1]\odot c^l_b[t-1]\Big)\\
= x^{l-1}_a[t]\odot c^{l}_b[t] + x^{l-1}_b[t]\odot c^l_a[t] + \frac{x^{l-1}_a[t]\odot x^{l-1}_b[t]}{t},
\end{gathered}
\end{equation}
Where \( c^l_a[t] \) and \( c^l_b[t] \) represent the membrane potentials at time step \( t \) for the two states, and \( x^{l-1}_a[t] \) and \( x^{l-1}_b[t] \) are the incremental inputs for these states at time step \( t \). The symbol \(\odot\) denotes element-wise multiplication. The output \( x^l[t] \) is the result of combining the two inputs over time and serves as the updated value at the time step \( t \).
We employ this mechanism to realize nonlinear operators that are not directly amenable to spiking conversion, including \texttt{LayerNorm}, \texttt{GELU}, and \texttt{Softmax} in attention, as well as multiplicative and matrix-multiplication composites within attention blocks. Meanwhile, the linear projections in attention are handled by the linear-layer conversion operation described earlier.
Building on differential coding, we introduce differential spiking neurons and perform unified differential conversion of linear and nonlinear operators in SigLIPv2, thereby obtaining a spiking vision encoder, SpikeSigLIP. It supports event-driven computation, improving computational efficiency and reducing redundant operations, while maintaining strong performance.
\subsection{Multimodal Spiking Large Language Model}
\label{sec:Spike-L}
Building on LLaMA-8B, we perform supervised fine-tuning by integrating real video, simulated data, auxiliary navigation, and VQA datasets to construct a foundational model for embodied navigation. Building on this, we integrate spike encoding mechanisms and Integrate-and-Fire neurons, using spike-driven computation and second-order optimization to construct a multimodal spiking large language model, thereby achieving low-power and efficient inference.
% We present Spike-L, which fuses visual and textual inputs into a unified token sequence and processes it with a spiking large language model.
% We encode visual and textual tokens using a generalized integrate-and-fire mechanism to generate temporally sparse representations and apply a channel saliency metric to allocate more spike steps to relevant channels, optimizing both computational and energy efficiency during inference.
\noindent\textbf{Unified Token Representations.}
We project the history frames, the current frame, and the text into a shared latent space and concatenate them into a unified token sequence, which is then processed by the spiking large language model.
\begin{equation}
{I}_t \;=\; \big[\,\mathbf{V}^{\mathrm{h}},\ \mathbf{V}^{\mathrm{c}},\ \mathbf{T}\,\big]
\in \mathbb{R}^{(196\times t + 196 + N_{\mathrm{text}})\times d},
% \vspace{-2mm}
\end{equation}
where $\mathbf{V}^{\mathrm{h}} \in \mathbb{R}^{(196\times t)\times d}$ are history visual tokens, $\mathbf{V}^{\mathrm{c}} \in \mathbb{R}^{196\times d}$ are current visual tokens, $\mathbf{T} \in \mathbb{R}^{N_{\mathrm{text}}\times d}$ are text tokens, $d$ is the shared hidden size, and $t$ is the number of history frames.
\noindent\textbf{Spiking Dynamics on Unified Tokens.}
To convert unified token representations into sparse spike tokens, we evolve each token through spiking neuron dynamics and aggregate responses across fine-grained time steps.
Each token receives a projected input and evolves with spike dynamics at fine time steps $t$:
\begin{equation}
\begin{aligned}
\tilde V_{i,t} &= \alpha\, V_{i,t-1} + ({I}_{i,t} W_{\mathrm{in}} + b) - \beta\, S_{i,t-1},\\
S_{i,t} &= \mathbf{1}\!\big(\tilde V_{i,t} \ge V_{th}), \quad
V_{i,t} = \tilde V_{i,t} - V_{th} S_{i,t},
\end{aligned}
% \vspace{-3mm}
\end{equation}
where ${I}_{i,t}$ is the $i$-th token, $W_{\mathrm{in}},b$ are input projection parameters, $V_{i,t}$ is the membrane potential, $S_{i,t}\!\in\!\{0,1\}$ is the spike, $\alpha\!\in\!(0,1)$ is the leak, $\beta\!\ge\!0$ models after the spike hyperpolarization, and $V_{th}$ is the threshold. To stabilize training,
% and reduce the effective temporal horizon,
we merge $L$ consecutive fine-grained steps into a multi-level spike token:
\begin{equation}
% \vspace{-3mm}
s_i[t']=\sum_{t=t'L}^{(t'+1)L-1} S_{i,t},\quad
t'=0,\ldots,T'-1, T'=\tfrac{T}{L}.
\end{equation}
where $L$ denotes the merge window, $T$ and $T'$ denote the numbers of fine-grained time steps before and after merging, respectively, and $s_i(t')\in\{0,\ldots,L\}$ is the spike count.
\noindent\textbf{Differential Temporal Sparsity Allocation.}
To evaluate channel importance both across modalities and within each modality, we propose a
differential temporal sparsity allocation mechanism. Specifically, informative channels are
assigned a longer spiking horizon, while less important ones are encoded with a single-step spike.
For a single modality, the spike-based reconstruction of the $c$-th channel at layer $l$ is formulated as:
\begin{equation}
{h}^{l}_{c}
=
\frac{V_{th}}{T_c}\sum_{t=1}^{T_c} p^{l}_{c}[t] + z^{l-1},
\label{eq:channel_spike_recon}
% \vspace{-3mm}
\end{equation}
where $p^{l}_{c}[t]\in\{0,1\}$ denotes the spiking activity of the channel $c$ at the fine-grained step $t$,
$T_c\in\{1,T'\}$ is the number of spiking steps assigned to the channel $c$, and $V_{th}$ is a
spiking threshold per-token. $z^{l-1}=\min({h}^{(l-1)})$ serves as a per-token zero-point shift, which allows spike-rate encoding for activations with negative values.
Considering channel importance across different modalities, we extend the channel-wise computation to the multimodal setting as follows:
\begin{equation}
\tilde{h}^{l}_c = \frac{V_{th}}{T'} \sum_{t=1}^{T'} p^{l}[t] \Big|_{p \in C_{\text{m}}} + \min(\tilde{h}^{l-1}_c),
\end{equation}
where \( C_{\text{m}} \) is the set of all channels within the current modality (e.g., visual or text). We employ second-order optimization to optimize the model, allowing for more precise weight adjustments and improving overall performance.
% and \( S_m \) is the set of significant channels in each modality \( m \in \mathcal{M} \).
% Next, we further extend this formula to allocate longer spiking steps to significant channels within each modality and consider cross-modal significance. The extended formulation, which accounts for channel importance both within each modality and across modalities, is given by:
% \begin{equation}
% \begin{split}
% \tilde{x}^{(k)} \approx & \frac{V_{\text{max}}}{L'} \sum_{t=1}^{L'} p^{(k)}_t \Big|_{p \in S_{\text{imp}}} + s^{(k)}(1) \cdot V_{\text{max}} \Big|_{p \in S_{\text{mod}}} \\
% &+ \min(\tilde{x}^{(k-1)}),
% \end{split}
% \end{equation}
% where \( S_{\text{imp}} \) represents the set of significant channels in the current modality, \( S_{\text{mod}} \) is the set of all channels within the current modality, and \( s^{(k)}(1) \) is the initial spike activity of the \( k \)-th channel. This extension allows for more detailed saliency weighting based on modality and channel importance.
The framework unifies visual and language representations into a spiking token stream using spike coding, driving event-based processing to achieve efficient computation and energy savings.
\subsection{Spiking Neural Network for Action Policy}
\label{sec:Spike-A}
We propose an action policy network based on Spiking Neural Networks. The input is encoded into spikes via population coding~\cite{deep} and processed by multiple fully connected SNN layers, where neuronal currents, membrane potentials, and spike events are updated iteratively over time. Finally, the spiking action embedding is decoded into a continuous action vector through a decoding module. During reinforcement learning, the network is trained using the Proximal Policy Optimization (PPO) algorithm to optimize its parameters.
\noindent\textbf{Encoding and Decoding Network.}
To encode continuous observation inputs into discrete spike outputs, we adopt population encoding. This approach transforms continuous observations into sparse and robust spike events, improving the stability and noise robustness of quadruped locomotion control. The encoding is given by the following formula:
\begin{equation}
\mathbf{A}_E(s)
= \Phi_{\mathrm{LoG}}(s;\boldsymbol{\mu},\sigma),
% \vspace{-1mm}
\end{equation}
where $\mathbf{A}_E(s)\in\mathbb{R}^{K}$ denotes the population responses for the scalar input $s$, and $\boldsymbol{\mu}=\{\mu_k\}_{k=1}^{K}$
and $\sigma$ are trainable receptive-field parameters initialized to cover the range of $s$.
\noindent\textbf{Deterministic Spike Encoding.}
The population response $\mathbf{A}_E(s)$ serves as the stimulus for an input neuron population, where
spikes are deterministically generated by simulating soft-reset integrate-and-fire (IF) dynamics.
For the $k$-th neuron, the membrane state and spike emission are updated as follows:
\begin{equation}
m_k^{t}=m_k^{t-1}+\lambda\,A_{E,k},
\end{equation}
\begin{equation}
y_k^{t}=\mathbf{1}\!\big(m_k^{t}\ge \theta\big),
\end{equation}
\begin{equation}
m_k^{t}\leftarrow m_k^{t}-(\theta-\varepsilon)\,y_k^{t},
\end{equation}
where $A_{E,k}$ denotes the stimulation strength of neuron $k$, $m_k^{t}$ is the membrane state at time $t$,
$y_k^{t}\in\{0,1\}$ is the emitted spike, $\lambda$ is an input scaling factor, $\theta$ is the firing threshold,
and $\varepsilon$ is a small constant controlling the soft reset.
In the decoding module, we accumulate the spike count for each action dimension over $T$ time steps
and compute the firing rate $fr(i)$ as the mean number of spikes per time step. The final action \(a_i\) for each output dimension is then generated by applying the learned decoding weight and bias. This can be expressed as:
\vspace{-5pt}
\begin{equation}
a_i = W_d^{(i)} \cdot \frac{1}{T} \sum_{t=1}^{T} m^t_K + b_d^{(i)}
\vspace{-5pt}
\end{equation}
Where $a_i$ is the action for the $i$-th output dimension, $W_d^{(i)}$ is the decoding weight for the $i$-th action dimension. $m^t_K $ represents the spike from the $i$-th output neuron in layer $K$ at time step $t$.
\noindent\textbf{Reinforcement Learning Training.}
We train Spike-A with PPO under an actor--critic framework. At timestep $t$, the actor maps the
observation $s_t$ to a policy $\pi_\theta(\cdot \mid s_t)=\mathcal{N}(\mu_t,\sigma_t)$ and samples
an action $a_t$, while the critic estimates the state value $V_\phi(s_t)$ to compute the advantage $\hat{A}_t$.
PPO stabilizes learning by clipping the policy ratio $r_t(\theta)$ to avoid excessively large updates.
We train the spiking action policy network using surrogate-gradient spatiotemporal backpropagation, accumulating
gradients over discrete timesteps $t=1,\dots,T$ and propagating them end-to-end through the spiking backbone
as well as the learnable population encoder and decoder.
Decoder parameters are updated independently for output populations $i=1,\dots,M$ corresponding to action
dimensions, and encoder parameters $[\mu^{(i)},\sigma^{(i)}]$ are updated independently for input populations
$i=1,\dots,N$. We perform periodic updates over a fixed temporal window of length $T$, resulting in stable
policies for continuous control.
% In training, we optimize the SNN with PPO by minimizing the clipped surrogate loss, where the actor maps observed states to actions and the critic provides value feedback to refine the policy.
% The policy ratio \( r_t(\theta) \) is the ratio of the action probabilities under the current policy \( \pi_\theta \) to those under the old policy \( \pi_{\theta_{\text{old}}} \), given by:
% \begin{equation}
% r_t(\theta) = \frac{\pi_\theta(a_t | s_t)}{\pi_{\theta_{\text{old}}}(a_t | s_t)}
% \end{equation}
% Where \( a_t \) is the action taken at time step \( t \).
% \( s_t \) is the observed state at time step \( t \).
% The advantage function \( \hat{A}_t \), computed by the critic network, indicates how much better the action \( a_t \) is compared to the baseline. It is given by:
% \begin{equation}
% \hat{A}_t = Q(s_t, a_t) - V(s_t)
% \end{equation}
% Where \( Q(s_t, a_t) \) is the action-value function, which represents the expected return from taking action \( a_t \) in state \( s_t \).
% - \( V(s_t) \) is the state-value function, which represents the expected return from state \( s_t \).
% The PPO objective is optimized using the surrogate loss function. This function balances maximizing the expected return with preventing excessively large updates by clipping the policy ratio. The PPO loss function is defined as:
% \begin{equation}
% L^{\text{PPO}}(\theta) = \mathbb{E}_t \left[ \min \left( r_t(\theta) \hat{A}_t, \text{clip}(r_t(\theta), 1 - \epsilon, 1 + \epsilon) \hat{A}_t \right) \right]
% \end{equation}
% Here, the clip function ensures that the policy update is bounded, preventing large deviations from the previous policy, which enhances stability.
We optimize the policy by maximizing the clipped surrogate objective. It can be formulated as follows:
\begin{equation}
J(\theta)=\mathbb{E}_t\big[
\min\!\big(
r_t(\theta)\hat{A}_t,\,
\mathrm{c}\big(r_t(\theta),1-\epsilon,1+\epsilon\big)\hat{A}_t
\big)
\big],
\end{equation}
where $r_t(\theta)=\pi_\theta(a_t\mid s_t)/\pi_{\theta_{\mathrm{old}}}(a_t\mid s_t)$ denotes the probability ratio between the current and previous policies, and $\hat{A}_t$ is the advantage estimate. The clipping operator bounds the effective update when $r_t(\theta)$ deviates beyond $[1-\epsilon,\,1+\epsilon]$, preventing overly large policy changes and improving optimization stability.
For continuous action spaces, the actor parameterizes the policy as a gaussian distribution with mean $\mu_t$ and standard deviation $\sigma_t$, from which the action $a_t$ is sampled, providing fine-grained continuous control.
% PPO trains the SNN action network by optimizing the surrogate loss function. The actor network generates actions based on the observed state, while the critic network estimates their value. The clipped surrogate loss ensures stable policy updates. Network parameters are updated using the PPO loss gradients, achieving an optimal policy that balances exploration and exploitation in continuous action spaces.
% The parameters of the SNN are updated using gradient descent. Specifically, for the output layer parameters \( W_i \) and \( b_i \) (\( i \in \{1, 2, \dots, M\} \)), the update rule is:
% \begin{equation}
% \nabla W_i L = \nabla a_i L \cdot W_i \cdot fr_i, \quad \nabla b_i L = \nabla a_i L \cdot W_i
% \end{equation}
% where \( \nabla a_i L \) is the gradient of the loss function \( L \) with respect to action \( a_i \), \( fr_i \) is the firing rate of population \( i \), and \( W_i \) is the connection weight.
% SNN parameters (weights and biases) are updated using spatiotemporal backpropagation. For each layer \( k \), the gradients are computed by accumulating over time steps \( T \) as follows:
% \begin{equation}
% \nabla W_k L = \sum_{t=1}^T o(t)^{(k-1)} \cdot \nabla c(t)^{(k)} L,
% \end{equation}
% \begin{equation}
% \quad \nabla b_k L = \sum_{t=1}^T \nabla c(t)^{(k)} L
% \end{equation}
% where \( o(t)^{(k-1)} \) represents the spikes from the previous layer, and \( \nabla c(t)^{(k)} L \) is the gradient of the loss function \( L \) with respect to the current layer's current \( c(t)^{(k)} \).
% For the input populations \( i \) (\( i \in \{1, 2, \dots, N\} \)), the update rule for encoding parameters (such as \( \mu_i \) and \( \sigma_i \)) is as follows:
% \begin{equation}
% \nabla \mu_i L = \sum_{t=1}^T \nabla o(t)(0)_i L \cdot A_E(i) \cdot \frac{s_i - \mu_i}{\sigma_i^2},
% \end{equation}
% \begin{equation}
% \quad \nabla \sigma_i L = \sum_{t=1}^T \nabla o(t)(0)_i L \cdot A_E(i) \cdot \frac{(s_i - \mu_i)^2}{\sigma_i^3}
% \end{equation}
% where \( \nabla o(t)(0)_i L \) is the gradient of the loss function \( L \) with respect to the input \( s_i \), \( A_E(i) \) is the output of the Fourier encoding, and \( \mu_i \) and \( \sigma_i \) are the mean and standard deviation of the input population.
% Every \( T \) time steps, all parameters are updated using these gradients. Through spatiotemporal backpropagation, gradients are propagated from the output layer to the input layer, ensuring that the weights and biases of each neuron population are updated appropriately.
% \( W_i \) and \( b_i \): The weights and biases of output population \( i \), controlling the strength of neuron connections and activation.
% \( a_i \): The action produced by output population \( i \), representing the decision made by the network.
% \( fr_i \): The firing rate of population \( i \), i.e., the average spike rate of the neurons in that population.
% \( o(t)^{(k-1)} \): The spike output of neurons from layer \( k-1 \) at time step \( t \).
% \( c(t)^{(k)} \): The current of neurons in layer \( k \) at time step \( t \).
% \( \nabla c(t)^{(k)} L \): The gradient of the loss function \( L \) with respect to current \( c(t)^{(k)} \).
% \( \mu_i \) and \( \sigma_i \): The mean and standard deviation of input population \( i \), which control the Fourier encoding process.
\begin{figure*}[t]
\centering \includegraphics[width=1\textwidth]{section/figure/compare_spike.pdf}
% \vspace{-6pt}
% \vspace{-18pt}
\caption{Comparison of resource consumption and performance between ANN-based and SNN-based architectures.}
\label{fig:compare_spike}
% \vspace{-12pt}
% \vspace{-12pt}
\end{figure*}
% \input{tables/vlm}
% \input{tables/benchmark}
% \input{tables/policy}
% \input{tables/quant}
\section{Experiments}
\subsection{Experimental Setups}
\noindent\textbf{Dataset \& Metric.}
We evaluate vision-and-language navigation on VLN-CE R2R val-unseen for unseen-scene generalization, VLN-CE RxR val-unseen for long-horizon instruction following, and VLN-CE-Isaac for end-to-end navigation under realistic dynamics and traversability constraints.
We use a unified benchmark that jointly evaluates navigation and low-level control. We evaluated VLN-CE R2R/RxR and VLN-CE-Isaac using a unified set of metrics, including NE, OS, SR, SPL, and nDTW, which capture goal-reaching accuracy, feasibility, success rate, path efficiency and trajectory fidelity, respectively. For low-level locomotion, we quantify command tracking and safety using linear and angular velocity tracking errors and the collision rate.
\input{tables/R2R}
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{section/figure/experiments.pdf}
% \vspace{-7mm}
\caption{\textbf{SNN action policy network ablations.} The top row compares different spike encoding kernels (Laplacian, Gaussian, Triangular, and IMQ) in terms of reward, linear velocity error, angular velocity error, and terrain difficulty over training steps. The bottom row reports reward curves under different hyperparameter settings, including the time step \(T\), population encoding size \(P\), Actor network dimensions \(A\), and Critic network dimensions \(C\).
}
\label{fig:spike_curve}
% \vspace{-6pt}
\end{figure*}
\input{tables/senor_abl}
\input{tables/RxR}
% \noindent\textbf{Training Details}
\subsection{Main Results}
\noindent\textbf{Navigation Performance on VLN-CE Benchmarks.}
We evaluate SpikeVLA on R2R-CE and RxR-CE to assess environmental generalization under strict energy constraints, and to test whether competitive navigation performance can be maintained in resource-constrained conditions.
On R2R-CE Val-Unseen, we evaluate SpikeVLA under a resource-constrained setting with RGB-only observations and no waypoint supervision. As shown in Table~\ref{tab:table1}, SpikeVLA achieves competitive navigation performance and slightly exceeds the strongest RGB baseline, NaVILA~\cite{navila}, on some metrics. SpikeVLA substantially reduces inference cost, lowering GPU memory usage from 16.1\,GB to 6.2\,GB and achieving an energy metric of $E{=}49.09J$, which is approximately 34\% of NaVILA.
A comparison with AO-Planner shows that, although AO-Planner~\cite{AO-Planner} uses additional sensor inputs such as panoramic views and depth, SpikeVLA achieves stronger core navigation metrics under the stricter RGB-only constraint. This suggests that the improvements come from more efficient representation learning and decision inference, rather than richer sensory inputs.
On RxR-CE Val-Unseen, the task presents higher language variability and a more challenging instruction distribution. Therefore, we evaluated SpikeVLA on this benchmark, providing both navigation performance and energy consumption. The current results suggest that SpikeVLA achieves performance comparable to strong baselines~\cite{Uni-navid, navila} under the same RGB-only, no-waypoint settings, while maintaining the energy-efficiency advantages observed in R2R-CE, as shown in Table~\ref{tab:table2}.
Comparison of resource consumption and performance between ANN-based and SNN-based architectures, shown in Fig.~\ref{fig:compare_spike}.
The left side shows a comparison of resource consumption between the different components of SpikeVLA and NaVILA~\cite{navila}, while the right side shows the performance comparison. It can be seen that SpikeVLA achieves comparable performance while using nearly one-third of the energy and computational resources of NaVILA. In general, SpikeVLA balances performance and efficiency, offering competitive navigation while reducing resource consumption for energy-constrained platforms.
% \begin{figure}[t]
% \centering \includegraphics[width=\columnwidth]{section/figure/test.png}
% \vspace{-1.6em}
% \caption{We compare ANN-based and SNN-based VLA architectures on R2R and RxR across standard navigation metrics.}
% \label{fig:compare_spike}
% % \vspace{-24pt}
% \vspace{-20pt}
% \end{figure}
\input{tables/policy}
\noindent\textbf{VLA Navigation Performance in Simulation.}
We evaluated SpikeVLA in the VLN-CE-Isaac simulator using the Unitree Go2 platform to assess its transferability to closed-loop embodied execution under realistic dynamics and sensor noise. As shown in Table~\ref{tab:table3}, it achieves a strong and stable navigation performance, maintaining high metric scores compared to NaVILA~\cite{navila}. These results indicate that the policy supports reliable task-oriented navigation behaviors when deployed in a closed-loop execution system.
% We evaluate SpikeVLA in the VLN-CE-Isaac simulator using the Unitree Go2 platform to assess its transferability to closed-loop embodied execution under realistic dynamics and sensor noise. As shown in Table~\ref{tab:table3}, both SpikeVLA-Blind, which relies only on proprioception, and SpikeVLA-Vision, which additionally uses LiDAR, achieve strong and stable navigation performance, maintaining high NE, OS, SR, and SPL scores. These results indicate that the policy supports reliable task-oriented navigation behaviors when deployed in a full closed-loop execution system.
% Moreover, the LiDAR-equipped model employs explicit local geometric cues, leading to lower navigation error, higher success rates, and improved path efficiency.
Overall, the simulation results validate the transition from offline benchmarks to embodied control, reinforcing SpikeVLA as a practical foundation for long-horizon deployment under energy constraints.
\input{tables/vla_abl}
\noindent\textbf{Action Policy Performance.}
We evaluate SpikeVLA for low-level policy execution, focusing on control accuracy, safety, and energy efficiency.
As shown in Table~\ref{tab:table4},
% SpikeVLA consistently achieves lower control errors compared to the ROA baseline, demonstrating more stable closed-loop tracking and enhanced safety.
SpikeVLA matches NaVILA~\cite{navila} overall, with only minor differences on a few metrics, while significantly reducing resource usage and energy consumption without sacrificing control quality. This advantage is primarily driven by the event-driven mechanism of spiking neural networks, which aligns more effectively with the deployment requirements of high-frequency closed-loop low-level control. We validated the control performance across various terrains, as shown in Fig~\ref{fig:control}.
\input{tables/quantized}
\noindent\textbf{Comparison with Quantized VLA.}
We further compared SpikeVLA with INT4-quantized VLA models, with the results summarized in Table~\ref{tab:quant}. The results show that although INT4 quantization reduces computational and storage costs by lowering numerical precision, it leads to clear performance degradation. In contrast, SpikeVLA effectively reduces computational and storage costs while maintaining high task accuracy, thereby demonstrating a better trade-off between performance and efficiency.
It is important to emphasize that SpikeVLA and quantized ANN models differ fundamentally in their underlying mechanisms. Quantization mainly reduces computational cost by lowering the numerical precision of parameters and activations, whereas SNNs rely on an event-driven spiking computation mechanism in which computation is triggered only when neurons are activated, thereby reducing redundant operations. Therefore, SpikeVLA does not simply trade accuracy for efficiency. instead, it achieves higher energy efficiency through a sparse, event-driven computational paradigm.
\begin{figure}[htp]
\centering \includegraphics[width=1\columnwidth]{section/figure/robotic.pdf}
% \vspace{-0.5em}
\caption{\textbf{Control performance across different terrain difficulties.} Left, middle, and right subfigures show rugged, sloped, and obstacle-filled terrains. The red points denote LiDAR point clouds, with green arrows representing the commanded velocity and blue arrows representing the actual velocity.}
\label{fig:control}
% \vspace{-22pt}
\end{figure}
\subsection{Ablation Study}
\noindent\textbf{Effect of Different SNN Modules.}
Fig.~\ref{fig:compare_spike} presents an ablation study that progressively converts the VLA components into spiking neural networks, where Spike-V, Spike-L, and Spike-A denote the spike vision encoder, multimodal spiking large language model, and spiking neural network for action Policy, respectively.
As spiking conversion is applied to additional components, navigation performance shows only a modest degradation, while compute and energy costs decrease monotonically with consistent additive benefits.
The vision and reasoning modules play a key role in significantly reducing energy consumption.
Extending the spiking control module brings additional efficiency improvements.
% Overall, these results indicate that progressive, module-wise spiking conversion offers a controllable balance between performance and efficiency.
\input{tables/ablation}
\noindent\textbf{Effect of Different Spike Encoding Mechanisms.}
We find that Laplacian-kernel spike encoding achieves better performance on legged-robot control tasks. It represents continuous observations with sparse and robust population responses, improving the stability and noise robustness of quadruped locomotion control. Compared to the Gaussian RBF, whose squared-distance decay results in highly localized activations, the Laplacian kernel's exponential $\ell_1$-distance decay preserves sensitivity to mid-range variations.
Moreover, compared to the hard truncation of the triangular kernel and the heavy-tailed long-range influence of the IMQ kernel, the Laplacian kernel offers a better trade-off between coverage and suppression of distant interference.
Because foot-ground contacts, impact events, and terrain perturbations introduce pronounced transient dynamics, a stable yet not excessively smooth encoding generates more consistent population-fire activity, which supports downstream policy optimization and enhances the final return.
Fig.~\ref{fig:spike_curve} presents a comparison of the performance of different encoding kernels and the effect of different parameters in the action policy network on reward.
\begin{table*}[t]
\centering
% \vspace{-2pt}
\caption{\textbf{Comparison with SOTA methods on the VLN-CE Benchmarks.} The table summarizes navigation performance metrics NE, OS, SR, and SPL, together with resource-efficiency metric, where Mem denotes GPU memory usage and Eng denotes energy consumption. NaVid and UniNaVid have the same model parameters, resulting in equal energy consumption.}
% \vspace{-0.5em}
\label{tab:r2r}
\resizebox{1\linewidth}{!}{
\begin{tabular}{l|c|c|cccc|ccc}
\toprule
\multirow{2}{*}{Method} & \multirow{2}{*}{Observation} & \multirow{2}{*}{Waypoint} & \multicolumn{4}{c|}{R2R Val-Unseen} & \multicolumn{3}{c}{Resource Efficiency}\\
& & & NE $\downarrow$ & OS $\uparrow$ & SR $\uparrow$ & SPL $\uparrow$ & Mem(MB) $\downarrow$ & Eng(J)$\downarrow$ & ACEs($10^{12}$)$\downarrow$\\
\midrule
CM2~\cite{cm2} & RGB. \& Depth \& Odom. & \ding{55} & 7.02 & 41.0 & 34.0 & 27.0 & - & - & - \\
WS-MGMap~\cite{WS-MGMap} & RGB. \& Depth \& Odom. & \ding{55} & 6.28 & 47.0 & 38.0 & 34.0 & - & - & - \\
CMA~\cite{CMA} & Pano. \& Depth \& Odom. & \checkmark & 6.20 & 52.0 & 41.0 & 36.0 & - & - & - \\
Sim2Sim~\cite{sim2sim} & Pano. \& Depth \& Odom. & \checkmark & 6.07 & 52.0 & 43.0 & 36.0 & - & - & - \\
GridMM~\cite{gridmm} & Pano. \& Depth \& Odom. & \checkmark & 5.11 & 61.0 & 49.0 & 41.0 & - & - & - \\
Ego$^{2}$-Map~\cite{ego-map} & Pano. \& Depth \& Odom. & \checkmark & 5.54 & 56.0 & 47.0 & 41.0 & - & - & - \\
DreamWalker~\cite{dreamwalker}& Pano. \& Depth \& Odom. & \checkmark & 5.53 & 59.0 & 49.0 & 44.0 & - & - & - \\
HAMT+ScaleVLN~\cite{hamt} & Pano. \& Depth \& Odom. & \checkmark & 4.80 & - & 55.0 & 51.0 & - & - & - \\
ETPNav~\cite{etpnav} & Pano. \& Depth \& Odom. & \checkmark & 4.71 & 65.0 & 57.0 & 49.0 & - & - & - \\
HNR~\cite{hnr} & Pano. \& Depth \& Odom. & \checkmark & 4.42 & 67.0 & 61.0 & 51.0 & - & - & - \\
% \midrule
AO-Planner~\cite{AO-Planner} & Pano. \& Depth & \ding{55} & 5.55 & 59.0 & 47.0 & 33.0 & - & - & - \\
NaVid~\cite{navid} & RGB. & \ding{55} & 5.47 & 49.0 & 37.0 & 35.0 & 14231.96 & 157.29 & 4376.68 \\
UniNaVid\cite{Uni-navid} & RGB. & \ding{55} & 5.58 & 53.3 & 47.0 & 42.7 & 14231.96 & 157.29 & 4376.68 \\
NaVILA\cite{navila} & RGB. & \ding{55} & 5.28 & 61.5 & 53.9 & 49.3 & 16119.98 & 141.25 & 3930.21 \\
MapNav\cite{mapnav} & RGB. & \ding{55} & 4.93 & 53.0 & 39.7 & 37.2 & - & - & - \\
\midrule
% \textbf{SpikeVLA(ours) L} & RGB.& \ding{55} &5.56&61.61&52.80&47.69&2.8GB&1.0\\
% \textbf{SpikeVLA(ours) V T8} & RGB.& \ding{55} &5.63&61.34&50.57&46.13&15.8GB&0.5815\\
% \textbf{SpikeVLA(ours) VL T8} & RGB.& \ding{55} &5.59&63.30&51.39&45.88&2.8GB&0.5815\\
% \textbf{SpikeVLA(ours) V T16} & RGB.& \ding{55} &5.17&63.13&54.70&49.85&15.8GB&0.9898\\
% \textbf{SpikeVLA(ours) VL T16} & RGB.& \ding{55} &5.38&63.40&53.29&47.90&2.8GB&0.9898\\
\textbf{SpikeVLA(ours)} & RGB. & \ding{55} & 5.38 & 63.4 & 53.3 & 47.9 & 6249.18 & 49.09 & 1196.16 \\
\bottomrule
\end{tabular}}
% \vspace{-12pt}
\label{tab:table1}
\end{table*}
\begin{table}[!t]
\centering
% \small
\setlength{\tabcolsep}{2pt}
\caption{\textbf{VLN-CE-Isaac evaluation results.} NaVILA result is reproduced under the same conditions. All experiments evaluated on 1,077 episodes in the VLN-CE-Isaac.}
% \vspace{-0.6em}
\resizebox{1\linewidth}{!}{
\begin{tabular}{l|cccc|ccc}
\toprule
& \multicolumn{4}{c|}{VLN-CE-Isaac} & \multicolumn{3}{c}{Resource Efficiency}\\
& NE $\downarrow$ & OS $\uparrow$ & SR $\uparrow$ & SPL $\uparrow$ & Mem(MB)$\downarrow$ & Eng(J)$\downarrow$ & ACEs($10^{12}$)$\downarrow$ \\
\midrule
% \multicolumn{11}{l} \\
% NaVILA-Blind* & \checkmark & & & 6.03* & 49.0* & 36.2* & 33.3* & & &\\
% NaVILA-Blind(reproduce) & \checkmark & & & & & & & & &\\
% NaVILA-Vision* & \checkmark & \checkmark & & \textbf{5.49*} & \textbf{58.7*} & \textbf{50.2*} & \textbf{45.5*}& & &\\
% NaVILA-Vision(reproduce) & \checkmark & & & 3.63 & 57.7 & 44.8 & 41.9 & & &\\
% SpikeVLA-Guassian-Blind &Unitree Go2& \checkmark & & 4.27 & 48.8 & 33.1 &29.2 \\
% NaVILA-paper & 5.49&58.7 & 50.2& 45.5& 16126.18 & 141.25 & 3930.21 \\
NaVILA-R & 6.29&52.1 & 36.5& 29.5 & 16126.18 & 141.25 & 3930.21 \\
SpikeVLA& 6.02 & 53.6 & 32.7& 28.5& 6251.53 & 44.23 & 1109.61\\
% \midrule
% \multicolumn{11}{l}{\bf\emph{Unitree H1}} \\
% NaVILA-Blind & \checkmark & & & 7.67* & 33.3* & 24.4* & 21.0*& & &\\
% NaVILA-Vision & \checkmark & & \checkmark & \textbf{5.86*}& \textbf{54.6*} & \textbf{45.3*} & \textbf{40.3*} & & &\\
\bottomrule
\end{tabular}}
% \vspace{-20pt}
\label{tab:table3}
\end{table}
\begin{table*}[t]
\centering
\caption{\textbf{Comparison with SOTA methods on the Val-Unseen split of RxR-CE.}
The table summarizes navigation performance metrics NE, SR, SPL and nDTW, together with resource-efficiency measures.
}
% \vspace{-6pt}
% \vspace{-0.5em}
\resizebox{1\linewidth}{!}{
\begin{tabular}{l|c|c|cccc|ccc}
\toprule
\multirow{2}{*}{Method} & \multirow{2}{*}{Observation} &
\multirow{2}{*}{Waypoint} & \multicolumn{4}{c|}{RxR Val-Unseen} & \multicolumn{3}{c}{Resource Efficiency}\\
% \cmidrule(lr){2-5} \cmidrule(lr){6-9} \cmidrule(lr){10-13} \cmidrule(lr){14-16}
& & & NE $\downarrow$ & SR $\uparrow$ & SPL $\uparrow$ & nDTW $\uparrow$ & Mem(MB) $\downarrow$ & Eng(J) $\downarrow$& ACEs($10^{12}$) $\downarrow$\\
\midrule
BEVBert~\cite{bevbert} & Pano. \& Depth \& Odom. & \checkmark & 4.00 & 68.5 & - & 69.6 &-&-&-\\
CMA~\cite{CMA} & Pano. \& Depth \& Odom. & \checkmark & 8.76 & 26.5 & 22.1 & 47.0 &-&-&-\\
ETPNav~\cite{etpnav} & Pano. \& Depth \& Odom. & \checkmark & 5.64 & 54.7 & 44.8 & 61.9 &-&-&-\\
HNR~\cite{hnr} & Pano. \& Depth \& Odom. & \checkmark & 5.50 & 56.3 & 46.7 & 63.5 &-&-&-\\
AO-Planner~\cite{AO-Planner} & Pano. \& Depth & \ding{55} & 7.06 & 43.3 & 30.5 & 50.1 &-&-&-\\
UniNaVid\cite{Uni-navid} & RGB. & \ding{55} & 6.24 &48.7 &40.9& & 14231.96 & 157.29 & 4376.68\\
NaVILA\cite{navila} & RGB. & \ding{55} & 6.12 & 52.3 & 46.1 & 61.0 &16119.98&141.25 &3930.21\\
\midrule
\textbf{SpikeVLA(ours)} & RGB.& \ding{55} & 6.20& 51.9& 45.3 & 60.4 & 6249.18 &49.09&1196.16\\
% \textbf{SpikeVLA(ours) V} & RGB.& \ding{55} & & & \\
% \textbf{SpikeVLA(ours) VL} & RGB.& \ding{55} & & & \\
\bottomrule
\end{tabular}}
% \vspace{-12pt}
\label{tab:table2}
\end{table*}
\begin{table}[!t]
\centering
% \small
\setlength{\tabcolsep}{2pt}
\caption{Closed-loop performance of the low-level policy.}
% \vspace{-0.6em}
\resizebox{1\linewidth}{!}{
\begin{tabular}{lccccc}
\toprule
\multirow{2}{*}{Method}& \multicolumn{1}{c}{Linear Vel.} & \multicolumn{1}{c}{Angular Vel.} & \multicolumn{3}{c}{Resource Efficiency} \\
\cmidrule(lr){4-6}
& Error $\downarrow$ & Error $\downarrow$ & Mem(MB)$\downarrow$ & Eng(\textmu J)$\downarrow$ & ACEs($10^{6}$)$\downarrow$ \\
\midrule
NaVILA & 0.23 & 0.38 & 1.20&5.80 &161.48\\
\midrule
% PopSAN & 0.43 & 0.30 &2.35&0.41&7.34\\
SpikeVLA & 0.42 & 0.29 &2.35&0.31&5.53\\
% laplacian popdim 5 T 5
\bottomrule
\end{tabular}}
% \vspace{-12pt}
% \vspace{-18pt}
\label{tab:table4}
\end{table}
% \begin{table}[!t]
% \centering
% % \small
% % \setlength{\tabcolsep}{2pt}
% \caption{Ablation results of SpikeVLA components.}
% % Spike-V denotes the spike neural network vision encoder, Spike-L represents the multimodal spiking large language model, and Spike-A indicates the spiking neural network for action policy.}
% \vspace{-0.5em}
% \resizebox{1\linewidth}{!}{
% \begin{tabular}{ccc|cccc|cc}
% \toprule
% \multirow{2}{*}{Spike-V} & \multirow{2}{*}{Spike-L} & \multirow{2}{*}{Spike-A} & \multicolumn{4}{c|}{VLN-CE-Isaac} & \multicolumn{2}{c}{Resource Efficiency}\\
% % \cmidrule(lr){3-5} \cmidrule(lr){6-9} \cmidrule(lr){10-12}
% & & & NE $\downarrow$ & OS $\uparrow$ & SR $\uparrow$ & SPL $\uparrow$ & M $\downarrow$ & E $\downarrow$ \\
% \midrule
% % \multicolumn{11}{l} \\
% % NaVILA-Blind* & \checkmark & & & 6.03* & 49.0* & 36.2* & 33.3* & & &\\
% % NaVILA-Blind(reproduce) & \checkmark & & & & & & & & &\\
% % NaVILA-Vision* & \checkmark & \checkmark & & \textbf{5.49*} & \textbf{58.7*} & \textbf{50.2*} & \textbf{45.5*}& & &\\
% % NaVILA-Vision(reproduce) & \checkmark & & & 3.63 & 57.7 & 44.8 & 41.9 & & &\\
% % & & & & & & \\
% & & \checkmark & & & \\
% &\checkmark &\checkmark & & & \\
% \checkmark &\checkmark & \checkmark & & & \\
% % \midrule
% % \multicolumn{11}{l}{\bf\emph{Unitree H1}} \\
% % NaVILA-Blind & \checkmark & & & 7.67* & 33.3* & 24.4* & 21.0*& & &\\
% % NaVILA-Vision & \checkmark & & \checkmark & \textbf{5.86*}& \textbf{54.6*} & \textbf{45.3*} & \textbf{40.3*} & & &\\
% \bottomrule
% \end{tabular}}
% % \vspace{-5pt}
% \vspace{-18pt}
% \label{tab:table5}
% \end{table}
\begin{table}[t]
\setlength{\tabcolsep}{1.5pt}
\centering
% \vspace{-2pt}
\caption{Comparison with ANN quantized model.}
% \vspace{-0.5em}
\resizebox{1\linewidth}{!}{
\begin{tabular}{l|cccc|ccc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{4}{c|}{R2R Val-Unseen} & \multicolumn{3}{c}{Resource Efficiency}\\
& NE $\downarrow$ & OS $\uparrow$ & SR $\uparrow$ & SPL $\uparrow$ & Mem(GB) $\downarrow$ & Eng(J)$\downarrow$ & ACEs($10^{12}$)$\downarrow$\\
\midrule
NaVILA (FP16) &5.28 &61.5 &53.9 &49.3& 15.7& 141.25 &3930.21 \\
NaVILA (INT4)& 5.66 &56.8& 48.2& 43.6 &8.6& 72.49& 982.55\\
\midrule
\textbf{SpikeVLA}& 5.38 &63.4 &53.3 &47.9& 6.1& 49.09& 1196.16\\
\bottomrule
\end{tabular}}
% \vspace{-12pt}
\label{tab:quant}
\end{table}
\begin{table}[!t]
\centering
% \fontsize{12}{12}\selectfont
\setlength{\tabcolsep}{1.5pt}
\caption{Rewards across different population encoders. MEL donates Mean Episode Length, where higher values indicate better survival and task persistence. }
% \vspace{-0.5em}
\resizebox{1\linewidth}{!}{
\begin{tabular}{lccccc} % 保持 5 列结构
\toprule
\multirow{2}{*}{Kernel Classes} & \multirow{2}{*}{Rewards} & \multirow{2}{*}{MEL$\uparrow$}&\multicolumn{3}{c}{Resource Efficiency} \\
\cmidrule(lr){4-6} % 对应 M, E, ACEs 这三列
& & & Mem(MB)$\downarrow$ & Eng(\textmu J)$\downarrow$ & ACEs($10^{6}$)$\downarrow$ \\
\midrule
ANN & 33.45&976.81 & 1.20 & 5.80 & 161.48 \\
Gaussian RBF kernel & 23.10& 973.11& 2.35 & 0.41 & 7.34 \\
Inverse Multiquadric kernel & 22.73 &939.35& 2.35 & 0.68 & 12.06 \\
Triangular kernel & 25.15 &966.29& 2.35 & 0.25 & 4.42 \\
Laplacian kernel & 26.72&983.94 & 2.35 & 0.31 & 5.53 \\
\bottomrule
\end{tabular}}
\vspace{-12pt} % 保持模板的间距微调
\label{tab:table6}
\end{table}
\section{Conclusion}
\label{sec:conclusion}
We present SpikeVLA, an event-driven paradigm for robotic tasks that reduces inference computation and energy consumption while maintaining the ability to execute complex instructions. This system involves a trade-off between performance and efficiency. Experiments across multiple VLN benchmarks and closed-loop simulations demonstrate robust generalization and stable performance, alongside reduced energy consumption. Its effectiveness has been validated through deployment on GPU platforms. The new SpikeVLA paradigm has a revolutionary impact on resource-constrained platforms, such as micro-robots and deep space exploration robots, driving the application of low-power, high-performance computing in extreme environments.
% It suggest that systematically integrating event-driven computation into VLA provides a transferable approach for deploying embodied intelligence on resource-constrained platforms, while improving the efficiency–performance trade-off.
\noindent\textbf{Limitation and Future Work.} While SpikeVLA has demonstrated strong performance in simulations and on GPU platforms, its energy efficiency and real-time performance on neuromorphic chips remain unverified. Further validation on neuromorphic hardware will be the next step in our work.
Moreover, we will focus on exploring SpikeVLA applications on energy-constrained platforms, such as micro-robots and robots for deep space exploration.
\newpage
\appendix
\onecolumn
\section{Dataset \& Metric Details}
\subsection{Dataset.}
To comprehensively evaluate our method for vision-and-language navigation, we conduct experiments on three benchmarks: R2R-CE, RxR-CE, and VLN-CE-Isaac. R2R-CE is a continuous-control VLN benchmark set in photorealistic, reconstructed indoor environments. We evaluate on the val-unseen split to assess generalization across unseen scenes. Under the same continuous-control framework, RxR-CE presents a more challenging long-horizon navigation benchmark, including room-to-room traversal. We evaluate on the val-unseen split to assess robustness with complex instructions and extended navigation horizons. VLN-CE-Isaac is a high-fidelity benchmark built on Isaac Sim, where robots navigate indoor environments, with end-to-end evaluation conducted under realistic dynamics and traversability constraints.
\subsection{Metric.}
We use a unified evaluation set that includes navigation performance, low-level control stability, and energy consumption. For R2R-CE, RxR-CE, and VLN-CE-Isaac, we provide the standard VLN metrics as follows.
\subsubsection{VLN metrics.}
\begin{itemize}
\item \textbf{Navigation Error (NE~$\downarrow$)}: the Euclidean distance between the agent's final position and the goal, reflecting goal-reaching accuracy.
\item \textbf{Oracle Success Rate (OS~$\uparrow$)}: whether the agent enters the success threshold at any point along the trajectory, indicating the feasibility of reaching the goal.
\item \textbf{Success Rate (SR~$\uparrow$)}: Whether the final stop position falls within the success threshold, assessing task completion.
\item \textbf{Success-weighted Path Length (SPL~$\uparrow$)}: A success-based measure of path efficiency that penalizes unnecessary deviations and repeated steps, indicating whether the agent reaches the goal efficiently.
\item \textbf{Normalized Dynamic Time Warping (nDTW~$\uparrow$)}: A trajectory-alignment metric that measures the similarity between the predicted and reference paths in terms of both spatial geometry and temporal alignment, showing trajectory fidelity.
\end{itemize}
\subsubsection{Low-level control Metric.}
For low-level control, we evaluate performance using the following metrics to assess command tracking accuracy and safety.
\begin{itemize}
\item \textbf{Linear Velocity Error~$\downarrow$}: the deviation between the executed and target linear velocities, reflecting command tracking accuracy and motion stability.
\item \textbf{Angular Velocity Error~$\downarrow$}: the deviation between the executed and target angular velocities, reflecting turning and heading control accuracy.
% \item \textbf{Collision Rate~$\downarrow$}: the frequency or proportion of collisions during execution, reflecting obstacle-avoidance safety and closed-loop robustness.
\end{itemize}
\subsubsection{Theoretical Energy Consumption Evaluation}
\label{sec:theoretical_energy}
To ensure reproducible and fair efficiency comparisons across different implementations and hardware platforms, we use theoretical energy consumption as our energy metric. This metric estimates inference energy based on the model's computational scale and spike sparsity. Consistent with previous classic SNN works~\cite{yin2021accurate,yao2023attention,horowitz20141}, we assume that all operations are implemented with 45nm technology~\cite{yin2021accurate}, where EMAC = 4.6pJ and EAC = 0.9pJ. This standardized approach is used to evaluate the energy efficiency of SpikeVLA in comparison to ANN baselines. SpikeVLA is being integrated with neuromorphic hardware and a quadruped robot to assess its performance in low-power systems, as mentioned in the future work section.
% \noindent\textbf{Synaptic Operations (SOPs)}
We decompose \textsc{SpikeVLA} into a series of layers or computational blocks. For the $l$-th
layer, the synaptic operations are defined as follows:
\begin{equation}
\mathrm{SOPs}_l = r_l \cdot T \cdot \mathrm{FLOPs}_l ,
\label{eq:sops}
\end{equation}
where $r_l$ denotes the average firing rate of the input spike train to the layer, $T$ is the number of discrete simulation time steps, and $\mathrm{FLOPs}_l$ is the corresponding dense compute of the layer, dominated by MAC operations.
% \noindent\textbf{Theoretical Energy of SpikeVLA}
We account for two types of elementary operations, multiply-and-accumulate (MAC) and spike-based accumulation (AC). The total end-to-end inference energy of SpikeVLA is computed as follows:
\begin{equation}
E_{\mathrm{SpikeVLA}}
=
E_{\mathrm{MAC}}\cdot \mathrm{FLOPs}_1
+
E_{\mathrm{AC}}\cdot \sum_{l=2}^{L}\mathrm{SOPs}_l ,
\label{eq:energy_spikevla}
\end{equation}
where $L$ is the number of layers (or blocks), and $\mathrm{FLOPs}_1$ denotes the dense compute of the first layer. The remaining layers are computed using their synaptic operations. In practice,
we estimate $\mathrm{FLOPs}_l$ and $r_l$ for each layer in Spike-V, Spike-L, and Spike-A, and sum
their energies to obtain the total inference energy.
% \noindent\textbf{Theoretical Energy of ANN Baselines}
For ANN baselines, inference is fully performed with dense floating-point operations. The theoretical energy is therefore given by the following formula:
\begin{equation}
E_{\mathrm{ANN}}=E_{\mathrm{MAC}}\cdot \mathrm{FLOPs}_{\mathrm{ANN}} ,
\label{eq:energy_ann}
\end{equation}
where $\mathrm{FLOPs}_{\mathrm{ANN}}$ represents the total FLOPs of the ANN model during inference. This definition aligns with the SpikeVLA framework to ensure fair and reproducible comparisons.
\subsubsection{ACE metric.}
ACE, the Arithmetic Computation Effort~\cite{pokebnn}, is designed to reflect the computational cost of neural network inference on idealized hardware, specifically hardware implemented with CMOS technology. It is defined as:
\begin{equation}
ACE = \sum_{i \in I, j \in J} n_{i,j} \cdot i \cdot j,
\end{equation}
where \(n_{i,j}\) denotes the number of multiply-accumulate operations (MACs) between \(i\)-bit and \(j\)-bit numbers, and \(I\) and \(J\) are sets of bitwidths used in the inference.
\section{Details of Spiking Neural Network Visual Encoder}
% \input{tables/siglip}
% (i) \texttt{MTH\_diff\_line} is inserted before convolution and linear projection layers to enable temporal expansion and merging, as well as spiking-aware feature modulation. (ii) \texttt{exp\_comp\_neuron} wraps LayerNorm, GELU, and Softmax to stabilize normalization and nonlinear transformations during multi-step spiking dynamics. and (iii) \texttt{AtNeuron} is introduced in the attention module to enable event-driven attention computation. This design maintains compatibility with pretrained weights while enhancing sparse activation and improving energy efficiency. As summarized in Table~\ref{tab:spike_siglipv2_ops}, we present the detailed architecture of the spiking visual encoder.
\input{tables/ablation_ann2snn}
We use SigLIP-V2 as the ViT backbone, maintaining its original architecture, which includes patch embedding, 27 transformer encoder blocks, and an attention pooling head.
To support event-driven temporal inference and reduce dense computation, we apply spiking wrappers to the main dense operators.
Table~\ref{tab:ann2sn} compares the final-layer features between SigLIPv2 and the SNN vision encoder after ImageNet calibration. As the time-step increases from $T$=2 to $T$=16, the SNN vision encoder shows a gradual decrease in mean squared error (MSE), suggesting that longer time-steps improve feature alignment. Meanwhile, the stability of the firing rate increases, leading to an increase in energy consumption.
Although the energy consumption of the SNN vision encoder increases with longer time-steps, it maintains a strong balance between accuracy and energy efficiency. Overall, it demonstrates robust performance in long-horizon inference tasks, making it particularly suitable for energy-constrained applications.
\section{Additional Experiments}
% \begin{figure*}[t]
% \centering
% \includegraphics[width=1\textwidth]{section/figure/action_ablation.pdf}
% \vspace{-1.6em}
% \caption{Comparison of different parameters of action policy.}
% \label{fig:abla_action}
% \vspace{-16pt}
% \end{figure*}
% \begin{figure*}[t]
% \centering
% \includegraphics[width=1\textwidth]{section/figure/energy.png}
% \vspace{-1.6em}
% \caption{Comparison of resource efficiency of NaVILA(ANN) and SpikeVLA.}
% \label{fig:annvssnn}
% \vspace{-6pt}
% \end{figure*}
% vision encoder随时间T的均方误差
% vision encoder的ANN2SNN转换和spike时间步的设置、LLM量化的消融实验
\subsection{Ablation of Different Modules in SpikeVLA}
\input{tables/ablation_vlt}
Table~\ref{tab:vlt} presents the ablation study results of the SpikeVLA model on the R2R-CE Val Unseen and RxR-CE Val Unseen datasets. The results show that applying Spike-L results in a slight decrease in navigation performance, but significantly enhances energy efficiency, particularly in terms of memory and energy consumption. In contrast, Spike-V improves path efficiency and success rate, while also boosting computational efficiency.
Overall, through SNN-based visual encoding and SNN-based language inference, SpikeVLA achieves a favorable balance between high performance and low energy consumption.
Compared to the ablated versions of the vision or language modules, the full SpikeVLA model demonstrates enhanced overall performance, highlighting its advantages in long-horizon navigation tasks.
\subsection{Ablation Study of the Action Policy Model}
\subsubsection{Ablation Study of Hyperparameter Time Step}
% action 消融
% \input{tables/ablation_action}
\begin{figure*}[h]
\centering
\includegraphics[width=1\textwidth]{section/figure/T_ablation.pdf}
\vspace{-2em} \caption{\textbf{Ablation study of the time step hyperparameter.} It shows the performance of different time step values (T = 2, 3, 5) across various metrics, including reward, linear velocity error, angular velocity error, and terrain levels.}
\label{fig:T_ablation}
% \vspace{-16pt}
\end{figure*}
\input{tables/ablation_action_T}
Table~\ref{tab:action_ablation_T} presents the ablation study results of SpikeVLA-A across different timesteps $T$. As the time-step increases from $T$=2 to $T$=5, tracking accuracy improves, with the linear velocity tracking error decreasing from 0.44 to 0.35, and the angular velocity tracking error decreasing from 0.55 to 0.45, respectively. It suggests that longer time-steps improve the model's precision in dynamic tasks.
However, in terms of resource efficiency, memory consumption remains constant at 2.35$MB$, while energy consumption and the ACE metric increase substantially with longer time-steps. Energy consumption increases from 0.07 $\mu J$ at T=2 to 0.31 $\mu J$ at T=5, while ACE rises from 1.24 to 5.53. It suggests that while longer time-steps improve accuracy, they also increase computational and energy costs, emphasizing the trade-off between accuracy and resource consumption when choosing the optimal time-step. Figure~\ref{fig:T_ablation} demonstrates the effect of time step on the model's performance and stability across training steps.
\subsubsection{Ablation Study of Hyperparameter Population Encoding Size}
\begin{figure*}[h]
\centering
\includegraphics[width=1\textwidth]{section/figure/P_ablation.pdf}
\vspace{-2em} \caption{\textbf{Ablation study of the population encoding size hyperparameter.} It shows the performance of different population encoding dimensions (P = 2, 3, 5) across different metrics, including reward, linear velocity error, angular velocity error, and terrain levels.}
\label{fig:P_ablation}
% \vspace{-16pt}
\end{figure*}
\input{tables/ablation_action_P}
Table~\ref{tab:action_ablation_P} presents the ablation study results of SpikeVLA with different population encoding dimensions $P$. As the population encoding dimension increases from $P$=2 to $P$=5, tracking accuracy improves significantly, with the linear velocity tracking error decreasing from 0.77 to 0.35, and the angular velocity tracking error decreasing from 0.59 to 0.47. It suggests that increasing the population encoding dimension improves model accuracy, particularly in dynamic tasks, by giving for more precise feature representations.
However, in terms of resource efficiency, memory and energy consumption increase significantly as the population encoding dimension grows. Memory consumption increases from 0.98$MB$ and energy consumption from 0.03$\mu J$ at $P$=2, to 2.35$MB$ and 0.10$\mu J$ at $P$=5, while ACE increases from 0.53 to 1.76. This suggests that although increasing the population encoding dimension improves accuracy, it also results in higher computational and energy costs. Therefore, a balance between accuracy and resource consumption must be carefully considered in practical applications. Figure~\ref{fig:P_ablation} demonstrates the effect of population encoding size on the model's performance and stability across training steps.
\subsubsection{Ablation of Actor Network Dimensions}
\begin{figure*}[h]
\centering
\includegraphics[width=1\textwidth]{section/figure/act_ablation_c.pdf}
\vspace{-2em} \caption{\textbf{Ablation of Actor Network Dimensions.} It shows the performance of different Actor network dimensions (A = [128, 128], A = [256, 128], A = [256, 256]) across various metrics, including reward, linear velocity error, angular velocity error, and terrain levels, evaluated over training steps.}
\label{fig:act_ablation}
% \vspace{-16pt}
\end{figure*}
\input{tables/ablation_action_A}
Table~\ref{tab:action_ablation_A} presents the ablation study results of SpikeVLA with different Actor network dimensions. As the Actor dimension increases from [128, 128] to [256, 128] and [256, 256], there is no significant improvement in accuracy, with a slight degradation in the angular velocity tracking error. The tracking error for linear velocity decreases from 0.35 to 0.34, while the angular velocity error increases from 0.47 to 0.49, suggesting that increasing the dimension does not significantly improve accuracy and may even cause some performance degradation.
In terms of resource efficiency, memory consumption, energy consumption, and ACE increase significantly as the dimension grows. Memory consumption increases from 2.35MB and energy consumption from 0.10$\mu J$ at [128, 128] to 4.48MB and 0.34$\mu J$ at [256, 256]. This suggests that while larger Actor network dimensions may provide some accuracy improvements, they also incur higher computational and energy costs, emphasizing the trade-off between accuracy and resource consumption in practical applications. Figure~\ref{fig:act_ablation} illustrates the impact of different Actor network dimensions on the model's performance across various metrics, including reward, linear velocity error, angular velocity error, and terrain levels.
\subsubsection{Ablation of Critic Network Dimensions}
\begin{figure*}[h]
\centering
\includegraphics[width=1\textwidth]{section/figure/cri_ablation_c.pdf}
\vspace{-2em} \caption{\textbf{Ablation of Critic Network Dimensions.} This figure shows the performance of different Critic network dimensions (C = [512, 256, 128], C = [512, 256, 256], C = [256, 256, 128]) across various metrics, including reward, linear velocity error, angular velocity error, and terrain levels, evaluated over training steps.}
\label{fig:cri_ablation}
% \vspace{-16pt}
\end{figure*}
\input{tables/ablation_action_C}
Table~\ref{tab:action_ablation_C} presents the ablation study results of SpikeVLA with different Critic network dimensions. As the dimension increases from [256, 256, 128] to [512, 256, 128], accuracy improves, especially in the linear velocity tracking error, which decreases from 0.44 to 0.35, while the angular velocity error decreases from 0.51 to 0.47. This suggests that moderately increasing the Critic network dimensions improves accuracy.
However, when the Critic network dimensions are further increased to [512, 512, 256], accuracy drops significantly. The linear velocity error increases to 0.49, and the angular velocity error rises to 0.54. It indicates that excessively large network dimensions can degrade performance, likely due to model overfitting, which negatively impacts generalization and accuracy.
In terms of resource efficiency, memory consumption, energy consumption, and ACE values increase substantially with larger dimensions. Memory consumption increases from 2.35$MB$ to 4.63$MB$, energy consumption rises from 0.10$\mu J$ to 0.15$\mu J$, and ACE increases from 1.76 to 2.96. Overall, in practical applications, it is crucial to carefully balance accuracy and resource consumption when selecting the Critic network dimensions. Figure~\ref{fig:cri_ablation} illustrates the effect of different Critic network dimensions on the model's performance across various metrics, including reward, linear velocity error, angular velocity error, and terrain levels.
% \section{Qualitative Results of SpikeVLA}
% \begin{figure*}[t]
% \centering
% \includegraphics[width=1\textwidth]{section/figure/vis_nav.png}
% \caption{Vision and language navigation.}
% \label{fig:vis_nav}
% \end{figure*}
% \begin{figure*}[t]
% \centering
% \includegraphics[width=1\textwidth]{section/figure/vis_nav.png}
% \caption{Vision and language navigation.}
% \label{fig:vis_nav1}
% \end{figure*}
% \section{Vis}
% \begin{figure*}[h]
% \centering
% \includegraphics[width=0.72\textwidth]{section/figure/ANNvsSNN2.pdf}
% \vspace{0em} \caption{\textbf{Ablation of Critic Network Dimensions.} This figure shows the performance of different Critic network dimensions (C = [512, 256, 128], C = [512, 256, 256], C = [256, 256, 128]) across various metrics, including reward, linear velocity error, angular velocity error, and terrain levels, evaluated over training steps.}
% \label{fig:ann2snn}
% % \vspace{16pt}
% \end{figure*}
% \begin{table}[!t]
% \centering
% % \small
% % \setlength{\tabcolsep}{9pt}
% \caption{Performance of ANN2SNN on SigLIP 2 for ImageNet Dataset.}
% \resizebox{0.42\linewidth}{!}{
% \begin{tabular}{cccccc}
% \toprule
% Time-step $T$ & RMSE & IFR & AFR & Energy ratio \\
% \midrule
% 2 &2.359375 &0.3336&&\\
% 4 &1.4541015625 &0.4899&&\\
% 6 &0.66455078125 &0.4248&&\\
% 8 &0.3974609375 &&&\\
% 16 &0.1483154296875 &&&\\
% \bottomrule
% \end{tabular}}
% \vspace{-5pt}
% \label{tab:ann2sn}
% \end{table}
% \begin{table}[!t]
% \centering
% \caption{Performance of ANN2SNN on SigLIP 2 for ImageNet Dataset.}
% \resizebox{0.42\linewidth}{!}{
% \begin{tabular}{ccccc}
% \toprule
% Time-step $T$ & RMSE & IFR & AFR & Energy \\
% \midrule
% 2 & 2.3594 & 0.3336 & 0.2285 & 0.0894 \\
% 4 & 1.4541 & 0.4899 & 0.3502 & 0.2740 \\
% 6 & 0.6646 & 0.4248 & 0.3850 & 0.4520 \\
% 8 & 0.3975 & 0.3441 & 0.3789 & 0.5930 \\
% 16 & 0.1483 & 0.2434 & 0.3259 & 1.0204 \\
% \bottomrule
% \end{tabular}}
% \vspace{-5pt}
% \label{tab:ann2sn}
% \end{table}
\begin{table*}[h]
\centering
\setlength{\tabcolsep}{16pt}
\caption{Comparison of final-layer features between ANN and SNN visual encoders after ImageNet calibration.}
\renewcommand{\arraystretch}{1}
\resizebox{0.65\linewidth}{!}{
\begin{tabular}{ccccc}
\toprule
Time-step $T$ & MSE & IFR & AFR & Energy (J) \\
\midrule
2 & 2.4648 & 0.3456 & 0.2372 & 0.3091 \\
4 & 1.4209 & 0.4744 & 0.3526 & 2.2458 \\
6 & 0.6577 & 0.3950 & 0.3777 & 4.4890 \\
8 & 0.3704 & 0.3176 & 0.3665 & 6.2860 \\
10 & 0.2456 & 0.2723 & 0.3497 & 7.7378 \\
12 & 0.1863 & 0.2384 & 0.3324 & 8.9972 \\
14 & 0.1498 & 0.2190 & 0.3169 & 10.1200 \\
16 & 0.1237 & 0.2072 & 0.3034 & 11.1500 \\
\bottomrule
\end{tabular}}
\vspace{-5pt}
\label{tab:ann2sn}
\end{table*}
\begin{table}[h]
\centering
\caption{\textbf{Ablation study on the Val-Unseen split of R2R-CE and RxR-CE across different modules of the SpikeVLA.}}
\resizebox{1\linewidth}{!}{
\begin{tabular}{ccccccccccccc}
\toprule
\multirow{2}{*}{} & \multirow{2}{*}{$T$} & \multicolumn{4}{c}{R2R-CE Val Unseen} & \multicolumn{4}{c}{RxR-CE Val Unseen} &\multicolumn{3}{c}{Resource Efficiency}\\
\cmidrule(lr){3-6}\cmidrule(lr){7-10}\cmidrule(lr){11-13}
& & NE $\downarrow$ & OS $\uparrow$ & SR $\uparrow$ & SPL $\uparrow$ & NE $\downarrow$ & SR $\uparrow$ & SPL $\uparrow$ & nDTW $\uparrow$ & Mem(MB)$\downarrow$ & Eng(J)$\downarrow$ & ACEs($10^{12}$)$\downarrow$ \\
\midrule
Navila &-&5.28& 61.5& 53.9& 49.3&6.12& 52.3 &46.1& 61.0&16119.98&141.25&3930.21\\
\midrule
SpikeVLA(w/o V) &-&5.56&61.61&52.80&47.69&6.34&50.61&44.35&59.55&\underline{6248.23}&60.64&1629.38\\
\midrule
\multirow{2}{*}{SpikeVLA(w/o L)}&8&5.63&61.34&50.57&46.13&6.63&48.16&41.86&58.61&16125.93& 124.84 &3410.44\\
&16&\textbf{5.17}&63.13&\textbf{54.70}&\textbf{49.85}&6.19&51.86&46.02&60.76&16125.93& 129.70& 3496.99\\
\midrule
\multirow{2}{*}{SpikeVLA}&8&5.59&\underline{63.30}&51.39&45.88&6.52&47.83&41.83&58.34&\textbf{6249.18}&
\textbf{44.23}&\textbf{1109.61}\\
&16&\underline{5.38}&\textbf{63.40}&\underline{53.29}&\underline{47.90}&6.20&51.92&45.25&60.35&6249.18&\underline{49.09}&\underline{1196.16}\\
\bottomrule
\end{tabular}}
\label{tab:vlt}
\end{table}
% \textbf{SpikeVLA(ours) L} & RGB.& \ding{55} &5.56&61.61&52.80&47.69&2.8GB&1.0\\
% \textbf{SpikeVLA(ours) V T8} & RGB.& \ding{55} &5.63&61.34&50.57&46.13&15.8GB&0.5815\\
% \textbf{SpikeVLA(ours) VL T8} & RGB.& \ding{55} &5.59&63.30&51.39&45.88&2.8GB&0.5815\\
% \textbf{SpikeVLA(ours) V T16} & RGB.& \ding{55} &5.17&63.13&54.70&49.85&15.8GB&0.9898\\
% \textbf{SpikeVLA(ours) VL T16} & RGB.& \ding{55} &5.38&63.40&53.29&47.90&2.8GB&0.9898\\
\begin{table}[h]
\centering
\caption{Ablation study of time step T of SNN action policy network. The baseline configuration is marked with $\ast$.}
\label{tab:action_ablation_T}
\renewcommand{\arraystretch}{1} % 增加行高，更显优雅
\resizebox{1\linewidth}{!}{
\begin{tabular}{clcccccccc}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multicolumn{4}{c}{\textbf{Configuration}} & \multicolumn{2}{c}{\textbf{Tracking Error}} & \multicolumn{3}{c}{\textbf{Resource Efficiency} } \\
\cmidrule(lr){2-5}
\cmidrule(lr){6-7} \cmidrule(lr){8-10}
& Time Step & Pop. Dim & Actor Dim.& Critic Dim. &Linear Vel.$\downarrow$ & Angular Vel.$\downarrow$ & Mem(MB)$\downarrow$ & Eng(\textmu J)$\downarrow$ & ACEs($10^{6}$)$\downarrow$ \\
\midrule
% --- T 的消融 ---
\multirow{3}{*}{Spike-A}
& $T=2$ & $P=5$ & [128, 128] & [512, 256, 128] & 0.44 & 0.55 & 2.35 & 0.07 &1.24\\
& $T=3^\ast$ & $P=5$ & [128, 128] & [512, 256, 128] & 0.35 & 0.47 & 2.35 & 0.10 &1.76\\
& $T=5$ & $P=5$ & [128, 128] & [512, 256, 128] & 0.35 & 0.45 & 2.35 & 0.31 &5.53\\
\bottomrule
\end{tabular}}
\end{table}
\begin{table}[h]
\centering
\caption{Ablation study of population encoding size P. The baseline configuration is marked with $\ast$.}
\label{tab:action_ablation_P}
\renewcommand{\arraystretch}{1} % 增加行高，更显优雅
\resizebox{1\linewidth}{!}{
\begin{tabular}{clcccccccc}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multicolumn{4}{c}{\textbf{Configuration}} & \multicolumn{2}{c}{\textbf{Tracking Error}} & \multicolumn{3}{c}{\textbf{Resource Efficiency} } \\
\cmidrule(lr){2-5}
\cmidrule(lr){6-7} \cmidrule(lr){8-10}
& Pop. Dim & Time Step & Actor Dim.& Critic Dim. &Linear Vel.$\downarrow$ & Angular Vel.$\downarrow$ & Mem(MB)$\downarrow$ & Eng(\textmu J)$\downarrow$ & ACEs($10^{6}$)$\downarrow$ \\
\midrule
% --- T 的消融 ---
\multirow{3}{*}{Spike-A}
& $P=2$ & $T=3$ & [128, 128] & [512, 256, 128] & 0.77 & 0.59 & 0.98 & 0.03 &0.53\\
& $P=3$ & $T=3$ & [128, 128] & [512, 256, 128] & 0.65 & 0.68 & 1.43 & 0.15 &2.62\\
& $P=5^\ast$ & $T=3$ & [128, 128] & [512, 256, 128] & 0.35 & 0.47 & 2.35 & 0.10 & 1.76\\
\bottomrule
\end{tabular}}
\end{table}
\begin{table}[h]
\centering
\caption{Ablation study of actor network dimensions. The baseline configuration is marked with $\ast$.}
\label{tab:action_ablation_A}
\renewcommand{\arraystretch}{1} % 增加行高，更显优雅
\resizebox{1\linewidth}{!}{
\begin{tabular}{clcccccccc}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multicolumn{4}{c}{\textbf{Configuration}} & \multicolumn{2}{c}{\textbf{Tracking Error}} & \multicolumn{3}{c}{\textbf{Resource Efficiency} } \\
\cmidrule(lr){2-5}
\cmidrule(lr){6-7} \cmidrule(lr){8-10}
& Actor Dim. & Time Step & Pop. Dim & Critic Dim. &Linear Vel.$\downarrow$ & Angular Vel.$\downarrow$ & Mem(MB)$\downarrow$ & Eng(\textmu J)$\downarrow$ & ACEs($10^{6}$)$\downarrow$ \\
\midrule
% --- T 的消融 ---
\multirow{3}{*}{Spike-A}
& [128, 128] & $T=3$ & $P=5$ & [512, 256, 128] & 0.35 & 0.47 & 2.35 & 0.10 &1.76\\
& [256, 128]$^\ast$ & $T=3$ & $P=5$ & [512, 256, 128] & 0.43 & 0.48 & 4.63 & 0.17 &2.96\\
& [256, 256] & $T=3$ & $P=5$ & [512, 256, 128] & 0.34 & 0.49 & 4.48 & 0.34 &6.14\\
\bottomrule
\end{tabular}}
\end{table}
\begin{table}[h]
\centering
\caption{Ablation study of critic network dimensions. The baseline configuration is marked with $\ast$.}
\vspace{-6pt}
\label{tab:action_ablation_C}
\renewcommand{\arraystretch}{1} % 增加行高，更显优雅
\resizebox{1\linewidth}{!}{
\begin{tabular}{clcccccccc}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multicolumn{4}{c}{\textbf{Configuration}} & \multicolumn{2}{c}{\textbf{Tracking Error}} & \multicolumn{3}{c}{\textbf{Resource Efficiency} } \\
\cmidrule(lr){2-5}
\cmidrule(lr){6-7} \cmidrule(lr){8-10}
& Critic Dim. & Time Step & Pop. Dim & Actor Dim. &Linear Vel.$\downarrow$ & Angular Vel.$\downarrow$ & Mem(MB)$\downarrow$ & Eng(\textmu J)$\downarrow$ & ACEs($10^{6}$)$\downarrow$ \\
\midrule
% --- T 的消融 ---
\multirow{3}{*}{Spike-A}
& [512, 256, 128]$^\ast$ & $T=3$ & $P=5$ & [128, 128] & 0.35 & 0.47 & 2.35 & 0.10 &1.76\\
& [512, 256, 256] & $T=3$ & $P=5$ & [128, 128] & 0.49 & 0.54 & 2.35 & 0.09 &1.58\\
& [256, 256, 128] & $T=3$ & $P=5$ & [128, 128] & 0.44 & 0.51 & 2.35 & 0.15 &2.72\\
\bottomrule
\end{tabular}}
\vspace{-6pt}
\end{table}