Paper deep dive

DS-VLA: A Dendritic-inspired Vision-Language-Action Model for Robust Action Control

Yaxing Lyu, Jingyi Li, Mingkun Xu, Yujie Wu

Year: 2026Venue: arXiv preprintArea: cs.ROType: PreprintEmbeddings: 45

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 9/29/2026, 1:06:26 PM

Summary

The paper introduces DS-VLA, a dendritic-inspired Vision-Language-Action model designed to improve robustness in closed-loop robotic manipulation under transient action perturbations. By incorporating dendritic spiking dynamics, sparse branch-specific integration, and neuron-wise inhibitory gating, DS-VLA selectively suppresses unreliable state updates while preserving task-relevant history. It achieves a 91.6% nominal success rate and retains 95.4% of its performance under perturbations, significantly outperforming baselines like OpenVLA-OFT, FAST, π0, and GR00T.

Entities (8)

DS-VLA · model · 100%FAST · model · 95%GR00T · model · 95%LIBERO · benchmark · 95%OpenVLA-OFT · model · 95%π 0 · model · 95%Dendritic Spiking Dynamics · mechanism · 90%Inhibitory Gate · component · 90%

Relation Signals (7)

DS-VLA → evaluatedon → LIBERO

confidence 95% · We evaluate DS-VLA on all four LIBERO suites

DS-VLA → usescomponent → Inhibitory Gate

confidence 95% · we introduce a neuron-wise inhibitory gate that adaptively regulates the admission of new multimodal evidence into dendritic states

DS-VLA → usesmechanism → Dendritic Spiking Dynamics

confidence 95% · DS-VLA equips action neurons with multiple sparsely connected dendritic branches, each featuring heterogeneous, learned decay factors.

DS-VLA → outperforms → GR00T

confidence 90% · DS-VLA achieves... 87.35% average perturbed success rate... GR00T... achieve 30.75%

DS-VLA → outperforms → FAST

confidence 90% · DS-VLA achieves... 87.35% average perturbed success rate... FAST... achieve 23.90%

DS-VLA → outperforms → OpenVLA-OFT

confidence 90% · DS-VLA achieves a 91.6% average nominal success rate and an 87.35% average perturbed success rate... OpenVLA-OFT... achieve 39.45%

DS-VLA → outperforms → π 0

confidence 90% · DS-VLA achieves... 87.35% average perturbed success rate... π0... achieve 28.55%

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Vision-language-action (VLA) models have achieved strong performance in language-conditioned manipulation, yet success under nominal evaluation does not necessarily translate into robust closed-loop behavior when executed actions are transiently corrupted. We introduce DS-VLA, a dendritic-inspired action architecture that incorporates dendritic spiking dynamics into VLA control to address this limitation. Specifically, to enable modularized feature processing and temporal information integration, DS-VLA equips action neurons with multiple sparsely connected dendritic branches, each featuring heterogeneous, learned decay factors. Furthermore, to suppress unreliable state updates while preserving task-relevant historical information, we introduce a neuron-wise inhibitory gate that adaptively regulates the admission of new multimodal evidence into dendritic states prior to somatic dynamics. We evaluate DS-VLA on all four LIBERO suites under both nominal rollouts and a unified closed-loop action-perturbation protocol. DS-VLA achieves a 91.6\% average nominal success rate and an 87.35\% average perturbed success rate, retaining 95.4\% of its nominal performance. Under the same reported perturbation setting, OpenVLA-OFT, FAST, $\pi_0$, and GR00T achieve 39.45\%, 23.90\%, 28.55\%, and 30.75\%, respectively. A controlled ablation isolates the contribution of neuron-wise shared inhibition, while analyses of neural dynamics and post-perturbation trajectories associate robust performance with selective evidence suppression and effective behavioral recovery. Together, these results demonstrate that integrating brain-inspired computational mechanisms offers a promising architectural prior for robust embodied intelligence beyond merely scaling vision-language backbones or generative action decoders.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

42,041 characters extracted from source content.

Expand or collapse full text

DS-VLA: A Dendritic-inspired Vision-Language-Action Model for Robust Action Control Yaxing Lyu Jingyi Li Mingkun Xu Yujie Wu Abstract Vision–language–action (VLA) models have achieved strong performance in language-conditioned manipulation, yet success under nominal evaluation does not necessarily translate into robust closed-loop behavior when executed actions are transiently corrupted. We introduce DS-VLA, a dendritic-inspired action architecture that incorporates dendritic spiking dynamics into VLA control to address this limitation. Specifically, to enable modularized feature processing and temporal information integration, DS-VLA equips action neurons with multiple sparsely connected dendritic branches, each featuring heterogeneous, learned decay factors. Furthermore, to suppress unreliable state updates while preserving task-relevant historical information, we introduce a neuron-wise inhibitory gate that adaptively regulates the admission of new multimodal evidence into dendritic states prior to somatic dynamics. We evaluate DS-VLA on all four LIBERO suites under both nominal rollouts and a unified closed-loop action-perturbation protocol. DS-VLA achieves a 91.6% average nominal success rate and an 87.35% average perturbed success rate, retaining 95.4% of its nominal performance. Under the same reported perturbation setting, OpenVLA-OFT, FAST, π0 _0, and GR00T achieve 39.45%, 23.90%, 28.55%, and 30.75%, respectively. A controlled ablation isolates the contribution of neuron-wise shared inhibition, while analyses of neural dynamics and post-perturbation trajectories associate robust performance with selective evidence suppression and effective behavioral recovery. Together, these results demonstrate that integrating brain-inspired computational mechanisms offers a promising architectural prior for robust embodied intelligence beyond merely scaling vision–language backbones or generative action decoders. 1Guangdong Institute of Intelligence Science and Technology 2Hong Kong Polytechnic University 3The University of Hong Kong 4University of Illinois Urbana-Champaign yu-jie.wu@polyu.edu.hk, lyuyaxing@connect.hku.hk Code — https://github.com/ASTAR123/DS-VLA Introduction Vision–language–action (VLA) models connect multimodal foundation models with embodied control by mapping visual observations and language instructions to robot actions (Zitkovich et al. 2023; Kim et al. 2024). In recent years, progress in this field has been driven primarily by larger-scale pretraining data, stronger multimodal backbones, and advances in action modeling, exemplified by the autoregressive frequency-space action tokenization of FAST (Pertsch et al. 2025), the parallel continuous action-chunk regression of OpenVLA-OFT (Kim et al. 2025), the flow-matching action expert of π0 _0 (Black et al. 2024), and the dual-system architecture of GR00T with a diffusion-transformer action module for real-time motor generation (Bjorck et al. 2025). Fig. 1 contrasts these four representative designs under a common vision–language foundation-model interface (Community 2026). These advances have improved manipulation performance under standard evaluation, but nominal task success does not guarantee closed-loop robustness. Action noise, actuator errors, unexpected contact, or incorrect gripper states can alter both robot and object states, causing subsequent observations to deviate from the intended trajectory. A robust policy must therefore preserve task intent, limit disruptive updates to its internal state, and adapt its actions to the changed environment. Figure 1: Representative VLA action-modeling paradigms under a shared multimodal interface. FAST autoregressively decodes frequency-space action tokens, OpenVLA-OFT predicts continuous action chunks in parallel, π0 _0 generates actions through flow matching, and GR00T couples semantic reasoning with real-time motor generation through a dual-system architecture. Biological neural systems offer useful principles for constructing such action dynamics. Conventional artificial neural networks typically abstract a neuron as a point-like computational unit. In contrast, biological pyramidal neurons integrate inputs from different pathways locally within dendritic compartments before combining branch responses at the soma (Hausser et al. 2000; Gasparini and Magee 2006). Dendritic branches exhibit diverse connectivity and temporal response properties, allowing a single neuron to organize information across input subspaces and timescales. In spiking neural networks, explicitly modeling temporal heterogeneity across dendritic compartments has likewise been shown to support multi-timescale feature integration and improve robustness and generalization on temporal tasks (Zheng et al. 2024). As illustrated in Fig. 2, our abstraction maps this biological compartmentalization to sparse branch-specific receptive fields, distinct dendritic states, and somatic integration. Input partitioning can restrict the spread of local disturbances, dendritic state can attenuate transient fluctuations, and heterogeneous branch dynamics can jointly represent rapid changes and persistent task context. Dendritic compartments are therefore not merely a structural extension of point neurons; their local memory, temporal filtering, and branch-wise integration provide a natural dynamical prior for robust action modeling. Figure 2: From interconnected biological neurons to computational dendritic modeling. Top: two neurons connected through an axon–dendrite synapse. Bottom: synaptic inputs are integrated by separate dendritic branches before converging at the somatic unit. Cortical inhibitory interneurons further regulate dendritic processing in an activity-dependent and compartment-specific manner (Palmer et al. 2012). Rather than uniformly silencing an entire network, inhibitory circuits can selectively modulate particular neurons or dendritic regions. Martinotti cells, for example, provide dendrite-targeting inhibition to excitatory neurons, whereas neurogliaform cells can induce longer-lasting local GABAergic modulation through volume transmission (Walker et al. 2016; Oláh et al. 2009). Physiological and computational evidence shows that dendritic inhibition can exert a disproportionately strong influence on dendritic excitability, suppress NMDA-receptor-dependent nonlinearities, and gate the electrogenesis underlying burst firing (Goetz et al. 2015; Lovett-Barron et al. 2012). These findings motivate controlling the current admitted into the dendritic state in DS-VLA. Its proposed roles in selecting action-relevant evidence and limiting perturbation-induced state updates are computational hypotheses evaluated empirically in this work. Motivated by these principles, we propose DS-VLA, a dendritic-inspired Vision–Language–Action model for robust action control that introduces dendritic compartmentalization, spiking dynamics, and inhibitory modulation into VLA action generation. Multiple dendritic branches partition and integrate incoming features with distinct temporal responses, while somatic spiking dynamics organize information across time for continuous action prediction. DS-VLA further employs a neuron-wise shared inhibitory gate: an independent, input-dependent inhibitory coefficient is generated for each neuron and shared by its two dendritic branches. It reduces interference from task-irrelevant or phase-inconsistent evidence during nominal execution and limits anomalous state updates under perturbations, allowing the action module to protect accumulated task information while remaining responsive to meaningful environmental changes. DS-VLA focuses on how information enters, updates, and persists in the action state, providing a brain-inspired framework for studying the relationship between neural dynamics and embodied control. Our main contributions are summarized as follows: • We propose DS-VLA, a dendritic spiking action architecture for continuous robot control that effectively integrates neural information processing mechanisms into VLA backbones, providing a novel solution for improving the closed-loop stability of VLA policies. • We formalize and integrate two neural-inspired mechanisms within the VLA action generation process: branch-specific dendritic integration for multi-timescale feature processing, and neuron-wise shared inhibitory gating for selectively regulating state updates. The proposed unified design supports both effective information processing during nominal tasks and stable state retention under environmental perturbations. • We establish a systematic evaluation framework covering standard tasks, multiple perturbation levels, and controlled architectural ablation to examine task performance, perturbation resilience, and the role of inhibitory gating. Crucially, DS-VLA retains over 95% of its nominal performance under standardized action perturbations, significantly outperforming the evaluated baselines. Related Work Vision–Language–Action Models Large language models (LLMs) established scalable sequence modeling and transferable linguistic reasoning from large text corpora (Llama 2023). Vision–language models (VLMs) extended this paradigm by aligning visual and textual representations (Radford et al. 2021) and connecting visual encoders with instruction-tuned language models for multimodal reasoning (Liu et al. 2023b). Recent VLM research further moves from passive observation toward process-level reasoning and active critique in robotic manipulation (Liu et al. 2026). VLA models further ground these representations in control: RT-2 casts robot actions as tokens within a VLM (Zitkovich et al. 2023), while OpenVLA provides an open pretrained VLA for manipulation (Kim et al. 2024). Subsequent work has diversified action generation through frequency-space autoregressive tokenization in FAST (Pertsch et al. 2025), parallel continuous action-chunk regression in OpenVLA-OFT (Kim et al. 2025), flow matching in π0 _0 (Black et al. 2024), and dual-system inference in GR00T (Bjorck et al. 2025). StarVLA organizes such components into a modular development framework (Community 2026). Complementarily, NeuroVLA introduces a system-level neuromorphic hierarchy that associates high-level planning, adaptive stabilization, and fast spiking execution with cortex-, cerebellum-, and spinal-cord-like modules (Guo et al. 2026). Closed-Loop Robustness in VLA Control Robot policies operate in a closed loop, where an erroneous action can alter subsequent robot, object, and sensory states. Conventional robustness techniques improve generalization by randomizing visual appearance or physical dynamics during training (Tobin et al. 2017; Peng et al. 2018), while rapid motor adaptation infers environment-dependent control variables online (Kumar et al. 2021). Generative controllers such as Diffusion Policy model multimodal action distributions and improve visuomotor behavior on contact-rich tasks (Chi et al. 2025), but expressive action generation alone does not guarantee recovery after execution deviates from the nominal trajectory. STRONG-VLA uses a perturbation curriculum followed by clean-data refinement to improve robustness to visual and linguistic corruptions (Xie et al. 2026). RobustVLA evaluates perturbations in both VLA inputs and actions, and improves action robustness through optimization against worst-case action noise (Zhang et al. 2025). Because an action disturbance changes the physical state encountered at the next control step, robustness also depends on how the policy adapts its subsequent actions to new observations. We study this closed-loop recovery process through action-head dynamics and task success over the resulting trajectory. Spiking Neural Networks, Dendritic Computation, and Inhibitory Modulation Spiking neural networks (SNNs) maintain state through membrane dynamics and communicate with discrete events; surrogate-gradient methods make these temporal models trainable with gradient-based optimization (Neftci et al. 2019). Most artificial spiking neurons nevertheless remain point-neuron abstractions. Biological dendrites perform branch-local integration with nonlinear and state-dependent responses (Hausser et al. 2000; Gasparini and Magee 2006), motivating computational models that distribute inputs and temporal constants across dendritic compartments. Temporal dendritic heterogeneity has been shown to improve multi-timescale learning and robustness in SNNs (Zheng et al. 2024). Dendritic processing is further shaped by spatially and temporally specific inhibition in vivo (Palmer et al. 2012); even sparse dendritic inhibitory input can exert a strong effect on local excitability (Goetz et al. 2015). Dendrite-targeting interneurons regulate branch responses, while neurogliaform-mediated GABAergic volume transmission provides longer-lasting local modulation (Walker et al. 2016; Oláh et al. 2009). These findings support input-dependent, branch-related regulation as a computational principle. Prior work has largely studied these mechanisms in neuroscience, whereas their computational modeling and practical use in embodied intelligence have yet to be systematically investigated. Method Figure 3: Overview of DS-VLA. Visual observations and language instructions are encoded by the vision–language backbone, and action queries are fused with robot-state features through GRU-based encoding and FiLM modulation. The resulting action features are processed by the dendritic spiking action model and decoded into continuous robot actions. The lower panel details the neuron-shared inhibitory DH-LIF dynamics: each somatic neuron assigns one input-dependent inhibitory coefficient to its two dendritic branches, whose temporally integrated currents subsequently converge at the LIF soma to generate spikes. Given visual observations I, a language instruction c, and a robot-state history S=shh=1HS=\s_h\_h=1^H, DS-VLA predicts a continuous action sequence A^=a^tt=1T A=\ a_t\_t=1^T with a^t∈R7 a_t∈ R^7. As shown in Figure 3, the VLM and action queries extract multimodal features, while a Gated Recurrent Unit (GRU) (Chung et al. 2014) and Gated Feature-wise Linear Modulation (FiLM) (Perez et al. 2018) inject robot-state information before the dendritic spiking action head. Within each DH-LIF layer, sparse dendritic branches retain independent temporal states, and one input-dependent inhibitory coefficient is shared by the two branches of each somatic neuron before their currents converge at the soma. The overall computation is summarized as Z=⁡(⁡(I,c)),X(r)=ℳ⁡(Z,S(r)),A^(r)=ℱDS(X(r)). array[]cZ= Q( V(I,c)),\\ X^(r)= M(Z,S^(r)),\\ A^(r)= F_ DS(X^(r)). array (1) where V is the VLM encoder, Q extracts action queries, ℳ M performs state-conditioned modulation, and ℱDS F_ DS is the dendritic spiking action head. The superscript r∈0,1r∈\0,1\ indexes the two prediction rounds, and X(r)X^(r) denotes the state-conditioned action feature at round r. The first prediction updates the robot-state tensor through U, S(1)=⁡(A^(0),S(0))S^(1)= U( A^(0),S^(0)), and the second produces the final sequence A^=A^(1) A= A^(1). Multimodal and State-Conditioned Features Selected VLM hidden states HlH_l are projected to the action dimension, Kl=HlWp+bpK_l=H_lW_p+b_p. Eight learnable queries then extract action-relevant information through successive cross-attention blocks: Ql′=Ql−1+MHA⁡(LN⁡(Ql−1),Kl,Kl),Ql=Ql′+MLP⁡(LN⁡(Ql′)). array[]cQ _l=Q_l-1+ MHA( LN(Q_l-1),K_l,K_l),\\ Q_l=Q _l+ MLP( LN(Q _l)). array (2) Here, l indexes the selected VLM layers and their corresponding cross-attention blocks; Q0∈RB×8×768Q_0∈ R^B× 8× 768 denotes the eight learned action queries, and QlQ_l is their updated representation at block l. MHA MHA, LN LN, and MLP MLP denote multi-head attention, layer normalization, and the feed-forward block, respectively. The projected VLM features KlK_l serve as both keys and values, and the final query representation is Z=QLZ=Q_L. The final representation Z∈RB×8×768Z∈ R^B× 8× 768 is modulated by the robot state. A two-layer unidirectional GRU encodes the state history into its final hidden state hSh_S, while the eight query features are mean-pooled as z¯=18∑q=18Zq z= 18 _q=1^8Z_q. Their projected features are pZ=ϕ⁡(WZLN(z¯)+bZ),pS=ϕ⁡(WSLN(hS)+bS),p_Z=φ(W_Z LN( z)+b_Z), p_S=φ(W_S LN(h_S)+b_S), (3) where B is the batch size, ϕφ is ReLU, and all W and b terms are learned affine-projection parameters. A state gate forms pZ′=pZ⊙σ⁡(WgpS+bg)p _Z=p_Z σ(W_gp_S+b_g), where σ is the sigmoid function and ⊙ denotes element-wise multiplication. Two MLPs then predict the FiLM scale γ and shift β from the concatenated feature [pZ′;pS][p _Z;p_S]. The modulated query is Xq=Zq⊙(1+γ)+β.X_q=Z_q (1+γ)+β. (4) This feature-fusion gate is distinct from the neuron-wise inhibitory gate defined below. Dendritic Spiking Action Head Sparse Dendritic Integration and Inhibitory Gating The action head contains three dendritic spiking layers. Each somatic neuron has two dendritic branches that receive complementary input subsets through a fixed binary mask. For neuron i, branch b, layer l, and internal step t, the branch current is Ii,b,tl=∑jMi,b,jlWi,b,jlxj,tl+ci,bl,I_i,b,t^l= _jM_i,b,j^lW_i,b,j^lx_j,t^l+c_i,b^l, (5) where j indexes input features, xj,tlx_j,t^l is the input to layer l, Mi,b,jl∈0,1M_i,b,j^l∈\0,1\ is the fixed connectivity mask, and Wi,b,jlW_i,b,j^l and ci,blc_i,b^l are the learned weight and bias. Each neuron independently computes an input-dependent inhibitory coefficient from the complete layer input: gi,tl=σ⁡(∑jWi,jg,lxj,tl+big,l),I~i,b,tl=(1−gi,tl)Ii,b,tl. array[]c g_i,t^l=σ\! ( _jW_i,j^g,lx_j,t^l+b_i^g,l ),\\ I_i,b,t^l=(1-g_i,t^l)I_i,b,t^l. array (6) Here, Wg,lW^g,l and bg,lb^g,l are the learned gate parameters. The same gi,tl∈(0,1)g_i,t^l∈(0,1) modulates both branches of neuron i, but different neurons retain independent coefficients. Thus, the gate controls how strongly each neuron admits new evidence while preserving the relative organization of its two branch currents. Larger g denotes stronger inhibition. Dendritic and Somatic Dynamics Each branch maintains a state with a learnable decay βi,bd,l=σ⁡(τi,bl) _i,b^d,l=σ( _i,b^l): di,b,tl=βi,bd,ldi,b,t−1l+(1−βi,bd,l)I~i,b,tl.d_i,b,t^l= _i,b^d,ld_i,b,t-1^l+(1- _i,b^d,l) I_i,b,t^l. (7) Branch-specific decay factors provide heterogeneous temporal responses. Importantly, inhibition affects only the incoming-current term and therefore limits anomalous updates without erasing stored dendritic state. The soma sums its two branch states and applies a bounded membrane skip. Let λil _i^l be the learnable LIF somatic decay and αl=0.2σ(ρl)α^l=0.2σ(ρ^l) the layer-wise skip coefficient. The pre-spike membrane potential is u~i,tl=λilui,t−1l+∑bdi,b,tl+αlui,t−1l. u_i,t^l= _i^lu_i,t-1^l+ _bd_i,b,t^l+α^lu_i,t-1^l. (8) where ui,t−1lu_i,t-1^l is the post-reset membrane state from the preceding internal step, ∑bdi,b,tl _bd_i,b,t^l aggregates the two dendritic states, and ρlρ^l is a learned skip logit. The decay λil _i^l is initialized in [0.95,0.99][0.95,0.99] and learned independently for each somatic neuron. Spike generation and subtractive reset follow zi,tl=ℋ⁡(u~i,tl−ϑil),ui,tl=u~i,tl−zi,tlϑil. array[]cz_i,t^l= H( u_i,t^l- _i^l),\\ u_i,t^l= u_i,t^l-z_i,t^l _i^l. array (9) where ℋ H is the Heaviside step function, ϑil _i^l is the learnable firing threshold, and zi,tl∈0,1z_i,t^l∈\0,1\ is the output spike. A fast-sigmoid surrogate derivative is used during backpropagation. Continuous Action Readout and Training DS-VLA stacks three such layers, with binary spikes passed between adjacent layers. Each layer contains 1536 neurons with two branches; the first receives the 768-dimensional modulated query. Continuous actions are decoded from the cumulative mean of the final-layer membrane state: mt=1t∑k=1tuk3,a^t=Womt+bo.m_t= 1t _k=1^tu_k^3, a_t=W_om_t+b_o. (10) where uk3u_k^3 is the final-layer membrane state, mtm_t is its cumulative temporal mean, and Wo,boW_o,b_o are the learned continuous-action readout parameters. This readout retains subthreshold and temporal information unavailable in a purely binary spike representation. Neural states propagate across query positions, while their gradients are detached every eight internal steps to implement truncated backpropagation through time without resetting the states. Figure 4: Successful and failed DS-VLA trajectories under Difficulty 1. Red dashed arrows denote injected action displacement. The model is optimized end to end with L1L_1 action regression, where da=7d_a=7: ℒaction=1BTda∑n=1B∑t=1T‖a^n,t−an,t‖1. L_ action= 1BTd_a _n=1^B _t=1^T \| a_n,t-a_n,t \|_1. (11) where n indexes the B batch elements, t indexes the T predicted actions, and an,ta_n,t is the ground-truth action. No auxiliary gate, firing-rate, or dendritic regularization loss is used. Experiments We evaluate DS-VLA on standard manipulation, action perturbations of increasing severity, and inhibitory-gating ablations. Success rate is the primary metric, complemented by successful-trajectory and action-head analyses. Experimental Setup Benchmark and Tasks We evaluate LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and Long, which emphasize spatial, object, goal, and long-horizon generalization (Liu et al. 2023a). Each suite contains ten tasks with 50 episodes each, yielding 2,000 episodes. Images are resized to 224×224224× 224, and the robot-state history contains 16 observations. Ten initial no-op steps stabilize the simulator and are excluded from the policy horizon. Maximum horizons are 220, 280, 300, and 520 for Spatial, Object, Goal, and Long; predicted actions are de-normalized to the Franka execution space. Baselines We compare with OpenVLA-OFT, FAST, π0 _0, and GR00T, competitive policies representing parallel continuous regression, autoregressive action tokens, flow matching, and dual-system diffusion control, respectively. Their distinct action dynamics complement the stateful dendritic and spiking dynamics of DS-VLA. All methods use the same task definitions, episode counts, success interface, and environment. The ungated variant isolates the contribution of inhibitory gating. Evaluation Metric Our primary metric is the task success rate SR=NsuccNall×100%,SR= N_succN_all× 100\%, (12) where NsuccN_succ and NallN_all denote successful and evaluated episodes. We aggregate tasks within each suite and macro-average the four suites. Action-Perturbation Protocol We perturb the executed 7-DoF actions. Difficulty 1 injects translation/rotation Gaussian noise (0.1/0.0020.1/0.002) during steps 10–19 and flips the gripper command. Difficulties 2 and 3 extend the perturbation to 25 and 35 steps, increase the noise to 0.010/0.0040.010/0.004 and 0.015/0.0080.015/0.008, overwrite 10 and 12 translation commands with [0.50,0,0][0.50,0,0] and [0.65,−0.40,0][0.65,-0.40,0], and force gripper closure for 18 and 25 steps, respectively. No perturbation indicator or hand-designed recovery signal is provided. Main Results DS-VLA achieves 91.6% average success and 98.6% on Object (Table 1). Although trailing OpenVLA-OFT, FAST, and π0 _0 on average, it remains competitive and outperforms GR00T, showing that dendritic spiking dynamics preserve strong nominal manipulation capability. Suite DS-VLA OFT FAST π0 _0 GR00T Goal 94.6 97.9 96.1 95.8 86.0 Object 98.6 98.4 97.2 98.8 92.0 Spatial 89.8 97.6 97.3 96.8 92.0 Long 83.4 94.5 90.2 85.2 76.0 Average 91.60 97.10 95.20 94.20 86.50 Table 1: Success rate (%) on standard LIBERO tasks. OFT denotes OpenVLA-OFT. Standard success rates primarily evaluate nominal state–action trajectories and do not reveal recovery from execution-induced deviations. Clean performance alone is therefore insufficient to characterize closed-loop robustness. Under Difficulty 1, DS-VLA achieves 87.35% average success and retains 95.4% of its clean performance (Table 2). Its degradation is only 4.25 points, compared with at least 55.75 for the baselines, and it exceeds the strongest perturbed baseline by 47.9 points. On Long, DS-VLA reaches 77.8% while all baselines remain at or below 1.2%. Thus, DS-VLA combines competitive nominal execution with markedly stronger recovery from action-induced deviations. Suite DS-VLA OFT FAST π0 _0 GR00T Goal 86.6 54.4 27.4 30.8 31.8 Object 95.4 21.8 13.8 13.8 30.4 Spatial 89.6 80.4 54.4 68.6 60.8 Long 77.8 1.2 0.0 1.0 0.0 Average 87.35 39.45 23.90 28.55 30.75 Table 2: Success rate (%) under Difficulty 1 action perturbations. OFT denotes OpenVLA-OFT. Figure 4 contrasts successful and failed responses on the same Goal task. The successful trajectory compensates after perturbation and completes placement; the failed trajectory does not recover before termination. This robustness is consistent with the encoded biological principles. Dendritic compartments integrate signals locally with heterogeneous temporal responses, while dendrite-targeting inhibition strongly and selectively regulates branch excitability and somatic output (Palmer et al. 2012; Goetz et al. 2015). DS-VLA abstracts these mechanisms through branch-specific states and an input-dependent inhibitory coefficient shared by each neuron’s two branches. The states preserve task-relevant information, while the gate limits abrupt updates without disrupting paired-branch organization. The clean and perturbed results therefore support combining dendritic temporal integration with neuron-wise inhibitory regulation. Robustness under Increasing Difficulty As disturbances become longer and more severe, DS-VLA’s average success decreases from 87.35% to 71.65% and 46.00% (Figure 5), confirming the substantial challenge posed by these settings. Figure 5: DS-VLA success rates (%) under increasing action-perturbation difficulty. For successful trajectories, we analyze completion horizon, action frequency, inference and loop latency, and normalized first- and second-order action differences (A and J), which are not metric Cartesian acceleration or jerk. As severity increases, the horizon grows from 145.4 to 175.4 steps, while action frequency, latency, and variation remain stable (Table 3). Recovery therefore relies primarily on additional closed-loop corrections rather than increased per-step computation or action magnitude. Level Hor.↓ Hz↑ Inf.↓ Loop↓ |A|↓|A| |J|↓|J| D1 145.4 6.56 122.6 151.2 .029 .040 D2 155.2 6.52 123.8 152.7 .027 .037 D3 175.4 6.73 119.9 147.8 .028 .039 Table 3: Successful DS-VLA trajectories under increasing perturbation difficulty. Hor. is completion horizon; Inf. and Loop are in milliseconds. This profile is consistent with compartmentalized dendritic dynamics: branch-specific states and heterogeneous decay provide multiple timescales, while inhibitory gating limits abnormal overwriting. DS-VLA can thus preserve task context across repeated corrections without increasingly large or irregular actions. These success-conditioned statistics describe recovery rather than unconditional efficiency. At Difficulty 3, forced actions may move the robot or object far from the nominal trajectory and create unrecoverable states. Dendritic dynamics and inhibitory gating correct moderate deviations but cannot resolve failures requiring a new task-level strategy. Stronger VLM failure recognition, recovery reasoning, and replanning should therefore complement the action head. Inhibitory-Gating Ablation Removing inhibitory gating reduces clean and perturbed averages by 18.0 and 20.9 percentage points, respectively (Figure 6). Because the ungated variant retains the same dendritic compartments and spiking dynamics, this gap isolates the additional value of input-dependent inhibitory regulation beyond dendritic temporal integration alone. The largest improvements occur on Long (+48.4+48.4 clean and +42.2+42.2 perturbed), indicating that selective state updating is particularly valuable when task information must persist over extended horizons. Perturbed Object is the sole exception, where the ungated variant is 0.8 percentage points higher. Figure 6: Success rates (%) for the inhibitory-gating ablation. Action-Head Efficiency We compare the DS-VLA, OpenVLA-OFT, and GR00T action heads. FAST is excluded because it autoregressively generates action tokens within the VLM decoder and therefore has no separable action head for a like-for-like comparison. We report dense operations and an ideal event-driven estimate that skips fixed mask zeros and replaces spike-driven MACs with accumulations. Arithmetic energy is estimated by Earith=CMAC(ϵmul+ϵadd)+CACCϵadd,E_ arith=C_ MAC( _ mul+ _ add)+C_ ACC _ add, (13) where CMACC_ MAC and CACCC_ ACC count the corresponding operations. Following the CLIF implementation (Huang et al. 2024) and Lemaire et al.’s framework (Lemaire et al. 2022), we use ϵmul=3.1 _ mul=3.1 pJ and ϵadd=0.1 _ add=0.1 pJ. This operation-level proxy assumes firing rates of 0.10 and 0.05 and is not measured GPU energy. Head Params Dense Eff. MAC/ACC Energy DS-VLA 17.738M 141.644M 56.930M/3.052M 0.182 OFT 42.004M 335.774M 335.774M/– 1.074 GR00T 100.865M 78.327G 78.327G/– 250.645 Table 4: Analytical action-head cost. Energy is estimated arithmetic energy (mJ); effective operations assume ideal event-driven execution. DS-VLA uses 57.8% fewer parameters and dense MACs than OFT. Under the ideal event-driven assumptions, its estimated arithmetic energy is 5.90× lower than OFT and 1,377.17× lower than ten-step GR00T, indicating substantial potential for sparse event-driven hardware. On an RTX 4090 at batch size one, DS-VLA draws 104.27 W on average, compared with 248.13 W for OFT and 109.64 W for GR00T. Spike-based communication and sparse dendritic connectivity may yield further event-driven hardware gains (Davies et al. 2018; Blouw et al. 2019), depending on spike sparsity, hardware mapping, and native deployment. Conclusion We presented DS-VLA, a brain-inspired VLA model that translates dendritic computation and inhibitory regulation into robot control. DS-VLA partitions input across branches with distinct states and temporal responses, then integrates their activity through a spiking soma. Motivated by dendrite-targeting cortical inhibition, its neuron-wise shared gate regulates current entering two branches while preserving their states and organization. Rather than simulating complete biophysics, DS-VLA abstracts dendritic integration, temporal memory, somatic spiking, and inhibitory control. Together, these mechanisms determine how evidence enters the action state, persists, and how disruptive updates are constrained. These biological priors benefit embodied control. DS-VLA achieves 91.6% average success on standard LIBERO tasks and 87.35% under Difficulty 1 action perturbations, retaining 95.4% of its nominal performance and outperforming the strongest perturbed baseline by 47.9 percentage points. Removing inhibitory gating reduces clean and perturbed performance by 18.0 and 20.9 percentage points, respectively, showing that adaptive inhibitory regulation provides substantial gains beyond dendritic temporal integration alone. The spiking, sparse dendritic architecture also offers a path toward efficient event-driven execution. We therefore view DS-VLA as a step toward uniting brain-inspired computing and embodied intelligence: biological mechanisms can become actionable principles for building more robust, adaptive, and potentially energy-efficient robots. This perspective treats neuroscience not merely as inspiration, but as a source of testable mechanisms for robust and efficient embodied action generation under perturbations. References Bjorck et al. (2025) J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y. Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: an open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734. Cited by: Introduction, Vision–Language–Action Models. Black et al. (2024) K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al. π0 _0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164. Cited by: Introduction, Vision–Language–Action Models. Blouw et al. (2019) P. Blouw, X. Choo, E. Hunsberger, and C. Eliasmith Benchmarking keyword spotting efficiency on neuromorphic hardware. In Proceedings of the 7th annual neuro-inspired computational elements workshop, p. 1–8. Cited by: Action-Head Efficiency. Chi et al. (2025) C. Chi, Z. Xu, S. Feng, E. Cousineau, Y. Du, B. Burchfiel, R. Tedrake, and S. Song Diffusion policy: visuomotor policy learning via action diffusion. The International Journal of Robotics Research 44 (10-11), p. 1684–1704. Cited by: Closed-Loop Robustness in VLA Control. Chung et al. (2014) J. Chung, C. Gulcehre, K. Cho, and Y. Bengio Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555. Cited by: Method. Community (2026) S. Community StarVLA: a lego-like codebase for vision-language-action model developing. arXiv preprint arXiv:2604.05014. Cited by: Introduction, Vision–Language–Action Models. Davies et al. (2018) M. Davies, N. Srinivasa, T. Lin, G. Chinya, Y. Cao, S. H. Choday, G. Dimou, P. Joshi, N. Imam, S. Jain, et al. Loihi: a neuromorphic manycore processor with on-chip learning. Ieee micro 38 (1), p. 82–99. Cited by: Action-Head Efficiency. Gasparini and Magee (2006) S. Gasparini and J. C. Magee State-dependent dendritic computation in hippocampal ca1 pyramidal neurons. Journal of Neuroscience 26 (7), p. 2088–2100. Cited by: Introduction, Spiking Neural Networks, Dendritic Computation, and Inhibitory Modulation. Goetz et al. (2015) L. Goetz, A. Roth, and M. Häusser Dendritic inhibitory synapses punch above their weight. Neuron 87 (3), p. 465–468. Cited by: Introduction, Spiking Neural Networks, Dendritic Computation, and Inhibitory Modulation, Main Results. Guo et al. (2026) W. Guo, H. Zhang, P. Li, T. Cai, Z. Chen, Y. Guo, X. He, Y. Yang, Y. Sun, and H. Xiong A brain-inspired embodied intelligence for fluid and fast reflexive robotics control. arXiv preprint arXiv:2601.14628. Cited by: Vision–Language–Action Models. Hausser et al. (2000) M. Hausser, N. Spruston, and G. J. Stuart Diversity and dynamics of dendritic signaling. Science 290 (5492), p. 739–744. Cited by: Introduction, Spiking Neural Networks, Dendritic Computation, and Inhibitory Modulation. Huang et al. (2024) Y. Huang, L. Xiaopeng, H. Ren, F. Haotian, Y. Zhou, L. Zunchang, B. Cheng, et al. Clif: complementary leaky integrate-and-fire neuron for spiking neural networks. In Forty-first International Conference on Machine Learning, Cited by: Action-Head Efficiency. Kim et al. (2025) M. J. Kim, C. Finn, and P. Liang Fine-tuning vision-language-action models: optimizing speed and success. arXiv preprint arXiv:2502.19645. Cited by: Introduction, Vision–Language–Action Models. Kim et al. (2024) M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: an open-source vision-language-action model. arXiv preprint arXiv:2406.09246. Cited by: Introduction, Vision–Language–Action Models. Kumar et al. (2021) A. Kumar, Z. Fu, D. Pathak, and J. Malik Rma: rapid motor adaptation for legged robots. arXiv preprint arXiv:2107.04034. Cited by: Closed-Loop Robustness in VLA Control. Lemaire et al. (2022) E. Lemaire, L. Cordone, A. Castagnetti, P. Novac, J. Courtois, and B. Miramond An analytical estimation of spiking neural networks energy efficiency. In International conference on neural information processing, p. 574–587. Cited by: Action-Head Efficiency. Liu et al. (2023a) B. Liu, Y. Zhu, C. Gao, Y. Feng, Q. Liu, Y. Zhu, and P. Stone Libero: benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems 36, p. 44776–44791. Cited by: Benchmark and Tasks. Liu et al. (2023b) H. Liu, C. Li, Q. Wu, and Y. J. Lee Visual instruction tuning. Advances in neural information processing systems 36, p. 34892–34916. Cited by: Vision–Language–Action Models. Liu et al. (2026) Y. Liu, Y. Lyu, D. Gao, Z. Liang, W. Tang, S. Mu, X. Yang, and Y. Mu From passive observer to active critic: reinforcement learning elicits process reasoning for robotic manipulation. arXiv preprint arXiv:2603.15600. Cited by: Vision–Language–Action Models. Llama (2023) H. T. Llama Llama: open and efficient foundation language models. arXiv preprint arXiv:2407.21783. Cited by: Vision–Language–Action Models. Lovett-Barron et al. (2012) M. Lovett-Barron, G. F. Turi, P. Kaifosh, P. H. Lee, F. Bolze, X. Sun, J. Nicoud, B. V. Zemelman, S. M. Sternson, and A. Losonczy Regulation of neuronal input transformations by tunable dendritic inhibition. Nature neuroscience 15 (3), p. 423–430. Cited by: Introduction. Neftci et al. (2019) E. O. Neftci, H. Mostafa, and F. Zenke Surrogate gradient learning in spiking neural networks. IEEE Signal Processing Magazine 36 (6), p. 51–63. Cited by: Spiking Neural Networks, Dendritic Computation, and Inhibitory Modulation. Oláh et al. (2009) S. Oláh, M. Füle, G. Komlósi, C. Varga, R. Báldi, P. Barzó, and G. Tamás Regulation of cortical microcircuits by unitary gaba-mediated volume transmission. Nature 461 (7268), p. 1278–1281. Cited by: Introduction, Spiking Neural Networks, Dendritic Computation, and Inhibitory Modulation. Palmer et al. (2012) L. Palmer, M. Murayama, and M. Larkum Inhibitory regulation of dendritic activity in vivo. Frontiers in Neural Circuits 6, p. 26. Cited by: Introduction, Spiking Neural Networks, Dendritic Computation, and Inhibitory Modulation, Main Results. Peng et al. (2018) X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel Sim-to-real transfer of robotic control with dynamics randomization. In 2018 IEEE international conference on robotics and automation (ICRA), p. 3803–3810. Cited by: Closed-Loop Robustness in VLA Control. Perez et al. (2018) E. Perez, F. Strub, H. De Vries, V. Dumoulin, and A. Courville Film: visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, Cited by: Method. Pertsch et al. (2025) K. Pertsch, K. Stachowicz, B. Ichter, D. Driess, S. Nair, Q. Vuong, O. Mees, C. Finn, and S. Levine Fast: efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747. Cited by: Introduction, Vision–Language–Action Models. Radford et al. (2021) A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, p. 8748–8763. Cited by: Vision–Language–Action Models. Tobin et al. (2017) J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), p. 23–30. Cited by: Closed-Loop Robustness in VLA Control. Walker et al. (2016) F. Walker, M. Möck, M. Feyerabend, J. Guy, R. J. Wagener, D. Schubert, J. F. Staiger, and M. Witte Parvalbumin-and vasoactive intestinal polypeptide-expressing neocortical interneurons impose differential inhibition on martinotti cells. Nature communications 7 (1), p. 13664. Cited by: Introduction, Spiking Neural Networks, Dendritic Computation, and Inhibitory Modulation. Xie et al. (2026) Y. Xie, Y. Yan, Y. Zhao, H. Wang, and Y. Jin STRONG-vla: decoupled robustness learning for vision-language-action models under multimodal perturbations. arXiv preprint arXiv:2604.10055. Cited by: Closed-Loop Robustness in VLA Control. Zhang et al. (2025) H. Zhang, S. Zhang, J. Jin, Q. Zeng, R. Li, and D. Wang Robustvla: robustness-aware reinforcement post-training for vision-language-action models. arXiv preprint arXiv:2511.01331. Cited by: Closed-Loop Robustness in VLA Control. Zheng et al. (2024) H. Zheng, Z. Zheng, R. Hu, B. Xiao, Y. Wu, F. Yu, X. Liu, G. Li, and L. Deng Temporal dendritic heterogeneity incorporated with spiking neural networks for learning multi-timescale dynamics. Nature Communications 15 (1), p. 277. Cited by: Introduction, Spiking Neural Networks, Dendritic Computation, and Inhibitory Modulation. Zitkovich et al. (2023) B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. Rt-2: vision-language-action models transfer web knowledge to robotic control. In Conference on Robot Learning, p. 2165–2183. Cited by: Introduction, Vision–Language–Action Models.

Instant research discovery