acceptodds
Under review as a conference paper at ICLR 2027

Spike-driven Vision–Language–Action Model

Abstract

Vision–language–action (VLA) models bridge multimodal understanding and robotic control, advancing the dominant paradigm for embodied intelligence. However, most existing models rely on large Transformers, whose latency and energy costs hinder deployment on resource-constrained platforms. Through sparse event-driven computation, spiking neural networks offer a promising paradigm for high-performance and energy-efficient computing. Here, we propose the first Spike-driven VLA framework enabling end-to-end direct training for robotic manipulation, which mainly comprises three core components. First, we develop spiking visual and instruction encoders for multimodal perception, encoding visual observations and language instructions into sparse, reliable spike representations for subsequent cross-modal fusion. Then, we introduce Multi-Winner Spike Fusion for instruction-guided scene understanding, using bidirectional top- winner-take-all spike routing to suppress background interference and yield fused memory. Finally, we propose a Spike Action Chunking Transformer that incorporates spiking cross-attention over the fused memory and the current robot state, enabling efficient end-to-end generation of continuous action chunks for robotic control. Extensive experiments on LIBERO and Meta-World demonstrate that Spike-driven VLA achieves competitive performance with fewer parameters and lower estimated inference energy than conventional VLA models. This work establishes a foundational framework for neuromorphic VLA modeling, paving the way for future advances in resource-efficient embodied intelligence.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Spiking Transformer, Spike Self-attention

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_XcHvOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a directly trained spiking VLA for robotic manipulation. It combines spiking visual and language encoders, bidirectional competitive fusion, and an action chunk decoder. Experiments on Meta World, LIBERO, and LIBERO Plus report competitive success rates with a compact model and low estimated arithmetic energy.

Critical strengths
  1. The compact architecture achieves promising manipulation performance. With approximately 149 million parameters, it reaches 92.4% success on LIBERO and 72.4% on Meta World, supporting the feasibility of directly trained spiking manipulation policies.

  2. The fusion mechanism has useful ablation support. Adding competitive routing improves average LIBERO success from 90.2% to 92.4%, with larger gains on Spatial and Long tasks.

Critical weaknesses
  1. The distinction from SpikeVLA is not sufficiently established experimentally. SpikeVLA already introduces spiking visual, language, and action components. The present paper emphasizes direct training rather than conversion and targets manipulation, but provides no direct comparison or controlled experiment isolating these differences. Its contribution should therefore be distinguished more precisely from the existing spiking VLA framework.

  2. The efficiency evidence does not establish practical deployment gains. Energy is estimated using arithmetic counts and reference operation costs, excluding memory access and non-synaptic overhead. There are no measured latency or hardware energy results, and the encoder ablations change architecture and pretraining alongside the computation type. Consequently, the benefits attributable specifically to spiking computation remain unclear.

Questions for authors
  1. What are the specific mechanistic differences from SpikeVLA in spike encoding, temporal integration, multimodal fusion, and action decoding? Beyond replacing conversion with direct training, which changes are essential, and what ablations support them?

  2. Could the authors compare against compact ANN policies using INT8 or INT4 quantization under comparable training and evaluation settings, reporting task success, memory, measured latency, and energy? This would clarify whether the proposed spiking mechanisms provide benefits beyond conventional quantization.

10 comments

We thank you for your constructive comments and suggestions, which have been valuable in improving the quality of our work. We provide point-by-point responses below.

Weakness 1 and Question 1: What are the specific mechanistic differences from SpikeVLA in spike encoding, temporal integration, multimodal fusion, and action decoding? Beyond replacing conversion with direct training, which changes are essential, and what ablations support them?

Response: We acknowledge SpikeVLA’s[1] pioneering contribution to integrating spiking neural networks into VLA models. Our method differs from SpikeVLA in three key aspects.

(1) Training paradigm. SpikeVLA combines ANN-to-SNN conversion for perception with PPO training for action control. These stages optimize representation reconstruction and control returns, respectively. However, policy updates do not jointly refine the perception module, limiting task-specific feature adaptation. In contrast, our unified action loss jointly optimizes visual encoding, cross-modal fusion, and action decoding, allowing action errors to directly guide visual feature learning.

(2) Inference latency. Spike-driven VLA directly optimizes action prediction within a fixed (T=4) simulation window, supporting cross-layer spike propagation and asynchronous inference. In SpikeVLA’s conversion-based approach, (T) governs layer-wise ANN activation reconstruction. Under layer-wise blocking execution, reconstruction across (L) layers can require approximately (T\times L) sequential simulation steps. Our end-to-end framework avoids waiting for complete layer-wise activation reconstruction, helping reduce synchronization overhead.

(2) Architecture and scale. SpikeVLA relies on an 8B language backbone for multimodal reasoning. Our method uses MWSF for sparse cross-modal fusion and SpikeACT for parallel action-chunk prediction, with approximately 0.15B parameters in the complete model.

As we have not obtained a complete public implementation or checkpoints of SpikeVLA, we cannot currently provide a controlled comparison on identical tasks. The key differences are summarized below.

AspectSpikeVLASpike-driven VLA (Ours)
TrainingStage-wise conversion + PPOEnd-to-end action-supervised training
Temporal inference$ T\times N$ (T=4)Direct spike propagation: T=4
Parameters>8B0.15B

*Here, NN denotes network depth.

Weakness 2 and Question 2: Could the authors compare against compact ANN policies using INT8 or INT4 quantization under comparable training and evaluation settings, reporting task success, memory, measured latency, and energy? This would clarify whether the proposed spiking mechanisms provide benefits beyond conventional quantization.

Response: Thank you for the suggestion. We have added comparisons with published quantized VLA policies[2] and architecture-matched experiments to examine whether the benefits of Spike-driven VLA extend beyond low-bit representation.

Table 1. LIBERO success rates (%) and reported memory footprints. Quantized baseline results are taken from QuantVLA.

MethodSettingMemory (GB) ↓SpatialObjectGoalLongAvg. ↑
π0.5FP164.2798.599.097.593.597.1
π0.5 + DuQuantW4A81.1786.097.571.550.076.3
GR00T N1.5FP162.0292.092.086.076.086.5
GR00T N1.5 + DuQuantW4A80.7466.070.068.076.070.0
Spike-driven VLA (Ours)Spiking0.85†95.095.496.682.492.4

Spike-driven VLA achieves higher LIBERO success than QuantVLA , with a lower reported memory. To investigate the contribution of spatio-temporal optimization beyond low-bit representation, we further evaluated a single-step variant (T=1T=1) and an architecture-matched W8A4 ANN.

Table 2. Ablation results on LIBERO.

VariantAvg. success rate (%) ↑
Spike-driven VLA (T=1T=1)78.6
Matched ANN, W8A486.2
Spike-driven VLA (full model)92.4

Reducing the simulation window to T=1T=1 decreases success by 13.8%, highlighting the contribution of temporal processing. The full model also exceeds the W8A4 ANN by 6.2%. These findings support the role of joint spatio-temporal optimization in the spiking domain, with performance benefits beyond the tested low-bit ANN representation.

Reference:

[1] Song R, Nie D, Teng S, et al. SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks[J]. arXiv preprint arXiv:2606.27807, 2026.

[2] Zhang J, Hsieh Y, Wan Z, et al. Quantvla: Scale-calibrated post-training quantization for vision-language-action models[J]. arXiv preprint arXiv:2602.20309, 2026.

Which tasks did you evaluate on a physical robot, and what metrics did you use? Could you report the number of trials and the success rate for each task, along with the corresponding baseline results where available?

Reviewer_3pvpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Spike-driven VLA, a vision-language-action framework based on spiking neural networks for robotic manipulation. It includes a spiking visual encoder, a spiking language encoder, a multimodal fusion module, and an action decoder. The method is evaluated on LIBERO, Meta-World, and LIBERO-Plus. The results show promising task performance and lower estimated energy consumption than several existing methods.

Critical strengths
  1. The paper studies an important problem. Efficient vision-language-action models are useful for robotic control, especially on resource-limited devices. The paper explores how spiking neural networks can reduce computational costs while maintaining good task performance.
  2. The paper proposes a complete spike-driven VLA framework. The framework includes visual encoding, language encoding, cross-modal fusion, and action generation. The MWSF module selects task-relevant visual information, while SpikeACT predicts multiple future actions in one pass.
  3. The method shows promising results on several benchmarks. The model achieves 92.4% average success on LIBERO and 72.4% on Meta-World. It also evaluates robustness on LIBERO-Plus without additional training. These results suggest that spiking neural networks have potential for language-guided robotic manipulation.
Critical weaknesses
  1. Questionable novelty over existing spiking VLA methods. The paper claims to introduce the first directly trained spike-driven VLA framework, but the difference from existing spiking VLA methods is not sufficiently clear. The authors should provide a detailed comparison and reproduce strong existing spiking VLA baselines under the same training and evaluation settings.
  2. The energy efficiency claims lack hardware validation. The reported energy is estimated using fixed operation costs from a 45-nm reference model. Memory access and other hardware overheads are excluded. The authors should measure actual energy consumption, inference latency, and memory usage on suitable hardware, and compare against ANN-based VLA models under consistent settings.
  3. The experimental comparisons are not controlled. The baselines use different architectures, pretrained models, and training settings. The contributions of MWSF and SpikeACT are also not fully isolated. The authors should reproduce strong ANN and SNN baselines under matched settings, conduct more component ablations, and evaluate different routing strategies and action chunk lengths.
  4. The generalization and deployment claims are not sufficiently supported. Although the paper evaluates LIBERO-Plus, performance varies substantially across perturbation types, with only 34.5% success under sensor noise. All evaluations are conducted in simulation. The authors should test additional manipulation benchmarks, more challenging distribution shifts, and real-robot tasks to demonstrate robustness and practical applicability.
  5. Statistical reliability and result consistency need improvement. The main results rely on 50 rollouts per task, without sufficient analysis of variation across training seeds. The paper also reports inconsistent overall success rates for LIBERO-Plus in the main text and Table 3. The authors should run multiple independent training seeds, increase evaluation trials, report confidence intervals, and verify all reported results.

I will increase my score if the authors can address these concerns with convincing evidence, especially through controlled baseline comparisons, a more complete efficiency evaluation, and stronger generalization experiments.

Questions for authors
  1. How does Spike-driven VLA differ from existing spiking VLA methods in terms of architecture, training, and performance? Can the authors provide a controlled comparison with a strong existing baseline?

  2. Can the authors provide more complete energy and latency measurements, including memory access and other inference overheads? Is the reported efficiency advantage maintained on actual hardware?

  3. Can the authors isolate the contributions of MWSF and SpikeACT through controlled ablations and comparisons with standard ANN modules?

  4. How well does the method perform under stronger distribution shifts and on real robotic platforms? In particular, can the authors improve or explain the low success rate under sensor noise?

  5. Can the authors report results across multiple training seeds, provide confidence intervals, and resolve the inconsistent LIBERO-Plus results?

4 comments

We sincerely thank you for your constructive suggestions, which have helped us strengthen the manuscript. We address each comment in detail below.

Weakness 1 and Question 1: How does Spike-driven VLA differ from existing spiking VLA methods in terms of architecture, training, and performance? Can the authors provide a controlled comparison with a strong existing baseline?

Response: We acknowledge SpikeVLA’s[1] pioneering contribution to integrating spiking neural networks into VLA models. Our method differs from SpikeVLA in three key aspects.

(1) Training paradigm. SpikeVLA combines ANN-to-SNN conversion for perception with PPO training for action control. These stages optimize representation reconstruction and control returns, respectively. However, policy updates do not jointly refine the perception module, limiting task-specific feature adaptation. In contrast, our unified action loss jointly optimizes visual encoding, cross-modal fusion, and action decoding, allowing action errors to directly guide visual feature learning.

(2) Inference latency. Spike-driven VLA directly optimizes action prediction within a fixed T=4T=4 simulation window, supporting cross-layer spike propagation and asynchronous inference. In SpikeVLA’s conversion-based approach, TT governs layer-wise ANN activation reconstruction. Under layer-wise blocking execution, reconstruction across LL layers can require approximately $ T \times L $ sequential simulation steps. Our end-to-end framework avoids waiting for complete layer-wise activation reconstruction, helping reduce synchronization overhead.

(3) Architecture and scale. SpikeVLA relies on an 8B language backbone for multimodal reasoning. Our method uses MWSF for sparse cross-modal fusion and SpikeACT for parallel action-chunk prediction, with approximately 0.15B parameters in the complete model.

As we have not obtained a complete public implementation or checkpoints of SpikeVLA, we cannot currently provide a controlled comparison on identical tasks. The key differences are summarized below.

AspectSpikeVLASpike-driven VLA (Ours)
TrainingStage-wise conversion + PPOEnd-to-end action-supervised training
Temporal inference$ T\times N $ (T=4)Direct spike propagation: T=4
Parameters>8B0.15B

Here, NN denotes network depth.

Weakness 2 and Question 2: Can the authors provide more complete energy and latency measurements, including memory access and other inference overheads? Is the reported efficiency advantage maintained on actual hardware?

Response: Thank you for the suggestion. We benchmarked Spike-driven VLA against TurboVLA, a compact ANN baseline, on an NVIDIA H100 80GB HBM3 GPU with batch size 1 and matched numerical precision, inputs, and action horizons. Each method was timed over 2,400 calls using recorded observations from 40 LIBERO tasks. End-to-end timing spans raw 256×256 images and state inputs to 12-step environment-action outputs, including preprocessing, data transfers, complete-policy computation, and action postprocessing. Amortized latency is the mean policy latency divided by the 12 output actions.

MethodSuccess rate (%) ↑Policy latency, mean ± SD (ms/call) ↓P95 latency (ms/call) ↓Amortized latency (ms/action) ↓Peak GPU memory (GiB) ↓
TurboVLA97.720.63 ± 2.4724.681.7190.93
Spike-driven VLA (Ours)92.45.62 ± 0.085.760.4680.79

Spike-driven VLA achieves a 3.67× speedup and 15.1% lower peak GPU memory usage. The timing includes full QK affinity construction, Top-k selection and indexing, membrane updates, and all internal spiking time steps already included in the end-to-end latency.

Weakness 3 and Question 3: Can the authors isolate the contributions of MWSF and SpikeACT through controlled ablations and comparisons with standard ANN modules?

Response: Thank you for the suggestion. We conduct a 2 × 2 controlled ablation by replacing MWSF and/or SpikeACT with the corresponding ANN fusion and action-chunking modules from our Spike-driven VLA. All variants use the same encoders, training data and budget, action horizons, and evaluation protocol. Module capacities are matched as closely as possible, with parameter counts reported below.

Fusion moduleAction decoderParams (M)LIBERO success (%) ↑
ANN fusionANN decoder148.594.8
ANN fusionSpikeACT148.594.1
MWSFANN decoder148.593.3
MWSFSpikeACT148.592.4

With all four variants matched at 148.5M parameters, the ANN/ANN baseline achieves 94.8% LIBERO success. Replacing only the action decoder with SpikeACT yields 94.1%, a decrease of 0.7 percentage points, while replacing only the fusion module with MWSF yields 93.3%, a decrease of 1.5 points. Both reductions are smaller than the 3.7-point improvement observed when replacing SVE with DINOv3-B in the separate encoder ablation. These results suggest that the current visual encoder remains a more prominent performance bottleneck than the fusion and action-decoding modules.

Weakness 4 and Question 4 : How well does the method perform under stronger distribution shifts and on real robotic platforms? In particular, can the authors improve or explain the low success rate under sensor noise?

Thank you for the suggestion. We evaluated Spike-driven VLA and similar TurboVLA[2] on an SO-101 robotic platform using language-conditioned single-object and sequential multi-object pick-and-place tasks. Under the reported runtime configurations on the same H100 GPU, our method achieves approximately 6.4× faster policy inference with lower peak GPU memory usage.

MethodSuccess rate (%) ↑Policy latency, mean ± SD (ms/call) ↓P95 latency (ms/call) ↓Amortized latency (ms/action) ↓Peak GPU memory (GiB) ↓
TurboVLA70%19.48 ± 2.7923.421.2180.92
Spike-driven VLA (Ours)65%3.05 ± 0.023.070.1900.76

Each inference call predicts 16 actions; amortized latency is the mean call latency divided by 16. Timing includes policy inference and host–device transfers but excludes network communication and client-side image preprocessing. Detailed task settings, deployment configurations, and evaluation protocols are provided in the appendix. The videos of the real-robot experiments to the supplementary material.

To investigate the low success rate under sensor noise, we conducted an additional ablation. Noise can obscure bowl boundaries and distort gripper–bowl distance estimates, challenging the current softmax-free SVE’s ability to preserve the fine-grained spatial cues required for grasping and placement.

Visual encoderSuccess under sensor noise (%) ↑
SVE (Ours, softmax-free)34.5
SVE (with softmax)43.7

Adding softmax to SVE increases success from 34.5% to 43.7%, a 9.2-percentage-point gain. This improvement highlights SVE’s spiking self-attention as a key target for enhancing fine-grained spatial representations and robustness to visual noise.

Weakness 5 and Question 5: Can the authors report results across multiple training seeds, provide confidence intervals, and resolve the inconsistent LIBERO-Plus results?

Response: Thank you for the suggestion. We have repeated training with three independent random seeds and revised Tables 3 and 10 using a consistent aggregation protocol. Individual seed results, together with the mean ± sample standard deviation across seeds, are reported below. All statistics are computed without intermediate rounding and presented to one decimal place, yielding an overall mean success rate of 55.2%.

Perturbation categorySeed 7Seed 123Seed 48Mean ± sample SD
Camera Viewpoints41.240.240.940.8 ± 0.5
Robot Initial States46.749.448.148.1 ± 1.3
Language Instructions48.749.348.949.0 ± 0.3
Light Conditions84.284.484.084.2 ± 0.2
Background Textures61.361.861.461.5 ± 0.3
Sensor Noise34.535.835.235.2 ± 0.7
Object Layout67.567.767.467.5 ± 0.1
Overall (macro-average)54.955.555.155.2 ± 0.3

Reference:

[1] Song R, Nie D, Teng S, et al. SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks[J]. arXiv preprint arXiv:2606.27807, 2026.

[2] Xie H, Yao C, Wu X, et al. TurboVLA: Real-time vision-language-action model at 32 Hz on an RTX 4090 with< 1 GB VRAM[J]. arXiv preprint arXiv:2607.27205, 2026.

Dear Reviewer,

As the discussion period is drawing to a close, we would be grateful if you could let us know whether our responses have addressed your concerns. We would welcome further discussion and the opportunity to clarify any remaining issues.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_ErYaOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper combines a spiking visual encoder, frozen spiking language encoder, bidirectional Top-k winner-take-all fusion, and an action-chunk decoder into a VLA policy. With approximately 0.15B parameters, it reports 72.4% success on Meta-World and 92.4% on LIBERO, alongside low operation-based estimates of inference arithmetic energy.

Critical strengths
  1. Extending directly trained spiking networks to language-conditioned continuous control broadens their application scope and offers a worthwhile direction for robot policies with limited computational resources.

  2. Evaluation spans two manipulation benchmarks and several distribution shifts, examining both task completion and robustness to environmental changes.

  3. Component ablations help clarify the roles of fusion and perceptual representations. The appendix also documents training, action execution, and the scope of theoretical energy accounting, providing a basis for further verification.

Critical weaknesses
  1. WTA in Eq. (9) outputs binary Top-k membership, whose ordinary derivative with respect to affinities is zero almost everywhere. Eq. (15) aggregates V through this mask, leaving dedicated Q/K projections and action queries without ordinary gradients through this routing path. Although the paper defines surrogate gradients for spiking thresholds, it does not specify WTA’s backward rule, leaving the central end-to-end learning mechanism without a reproducible training definition.

  2. Appendix D estimates arithmetic energy per policy forward pass and explicitly excludes memory access and non-synaptic overhead. Full QK affinity construction, Top-k selection, and membrane updates still incur execution costs, yet complete-policy latency or measured energy is not reported, nor is a compact ANN comparison using consistent per-action accounting. These results support potential arithmetic savings but do not establish a practical deployment advantage.

  3. The encoder substitutions in Table 5 change both model capacity and pretrained representations while retaining spiking fusion and action decoding. They reveal the influence of perception, but do not isolate spiking computation under matched capacity, data, and decoding. Without such a continuous-network control, the performance–cost trade-off remains difficult to attribute to the spiking mechanism.

  4. Table 1 imports SmolVLA results [1] from 10 trials per task, versus 50 here, so the 4.2-point gain needs confirmation under matched initializations. On LIBERO-Plus, Table 3 reports a 54.9% category macro average, whereas Table 10 supports 53.2% from pooled counts. Both are reproducible calculations, but explicit aggregation labels and consistent baseline evaluation protocols are needed to interpret these comparisons.

[1] Mustafa Shukor et al. SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics. arXiv:2506.01844v1, 2025.

Questions for authors
  1. What are the exact WTA forward and backward rules, including directional k values and implementation or logs demonstrating nonzero training gradients for Q/K projections and action queries?
  2. Under matched hardware, precision, inputs, and action horizon, what are the measured complete-policy latency, Top-k overhead, and per-action cost relative to a compact ANN?
  3. Against a continuous network with matched pretraining data, capacity, training budget, and action decoding, what performance and cost benefits does spiking computation provide?
  4. Can you verify the Meta-World performance gap and its stability under matched initializations, and clarify the category-macro versus pooled success-rate conventions for all LIBERO-Plus methods?
4 comments

Thank you for the constructive comments and suggestions, which have been valuable in improving the quality of our work. We provide point-by-point responses below.

Weakness 1 and Question 1: WTA Backward Definition and Gradient Flow. Binary Top-k routing has zero ordinary derivatives almost everywhere. The manuscript does not specify its backward rule or demonstrate gradient propagation to the dedicated Q/K projections and action queries.

Response: We thank the reviewer for highlighting this omission. Our implementation defines a hard-forward, soft-backward rule through a custom autograd operator. The forward pass generates a binary Top-k mask HH and aggregates values as C=HVC=HV. During backpropagation, we use the full Jacobian of a masked softmax proxy:

A=QK⊤,P=MaskedSoftmax⁡(A/β),β=dhτs,A=QK^\top,\qquad P=\operatorname{MaskedSoftmax}(A/\beta),\qquad \beta=\sqrt{d_h}\tau_s,

where dhd_h is the attention-head dimension and τs>0\tau_s>0 is the surrogate-temperature hyperparameter. Given the upstream gradient G=∂L/∂HG=\partial\mathcal L/\partial H, the backward rule is

∂L∂Aij^=Pijβ(Gij−∑rPirGir).\widehat{\frac{\partial\mathcal L}{\partial A_{ij}}}=\frac{P_{ij}}{\beta}\left(G_{ij}-\sum_r P_{ir}G_{ir}\right).

Together with the surrogate gradients of spiking neurons, this rule provides a gradient path to the dedicated Q/K projections. Softmax is used exclusively to compute surrogate gradients during backpropagation; both training-time forward computation and inference retain binary Top-k routing. Following your suggestion, we monitored the gradient norms of the dedicated Q/K projection parameters during training. As summarized below, both parameter groups exhibited nonzero gradient norms in every monitored mini-batch (100%), providing empirical evidence of consistent gradient propagation to both projections across the monitored training batches.

Parameter groupMedian gradient normBatches with nonzero gradients (%)
Dedicated Q projections2.66e-05100%
Dedicated K projections2.78e-05100%

Weakness 2 and Question 2: The arithmetic-energy estimates exclude memory access and other execution costs. Practical deployment benefits require complete-policy measurements and a compact ANN comparison under consistent accounting.

Response: We thank the reviewer for this suggestion. We benchmarked Spike-driven VLA against TurboVLA [1], a compact ANN baseline, on an NVIDIA H100 80GB HBM3 GPU with batch size 1 and matched numerical precision, inputs, and action horizons. Each method was evaluated over 2,400 policy calls using recorded observations from 40 LIBERO tasks. End-to-end latency was measured from raw 256×256 images and robot states to 12-action outputs, including preprocessing, data transfers, complete policy computation, and postprocessing. These measurements account for full QK affinity construction, Top-k selection and indexing, membrane updates, and all internal spiking time steps.

MethodSuccess rate (%) ↑Policy latency, mean ± SD (ms/call) ↓P95 latency (ms/call) ↓Amortized latency (ms/action) ↓Peak GPU memory (GiB) ↓
TurboVLA97.720.63 ± 2.4724.681.7190.93
Spike-driven VLA (Ours)92.45.62 ± 0.085.760.4680.79

Spike-driven VLA achieves a 3.67× inference speedup and 15.1% lower peak GPU memory usage.

To further assess real-robot performance, we deployed both policies on an SO-101 platform for language-conditioned single-object and sequential multi-object pick-and-place tasks, with policy inference running on the same H100 GPU. Under the reported runtime configurations, Spike-driven VLA achieves approximately 6.4× faster policy inference and lower peak GPU memory usage.

MethodSuccess rate (%) ↑Policy latency, mean ± SD (ms/call) ↓P95 latency (ms/call) ↓Amortized latency (ms/action) ↓Peak GPU memory (GiB) ↓
TurboVLA7019.48 ± 2.7923.421.2180.92
Spike-driven VLA (Ours)653.05 ± 0.023.070.1910.76

For SO-101, each policy call predicts 16 actions. Timing includes policy inference and host–device transfers but excludes network communication and client-side image preprocessing. In both tables, amortized latency is the mean policy latency divided by the number of predicted actions. Detailed task settings, deployment configurations, and evaluation protocols are provided in Section H.1 of the revised manuscript (highlighted in blue), and real-robot demonstration videos are included in the supplementary material.

Weakness 3: Isolating the Contribution of Spiking Computation. The encoder substitutions in Table 5 change both capacity and pretrained representations. They do not isolate the performance–cost contribution of spiking computation.

Response: Thank you for the suggestion. We agree that the original DINOv3/SVE comparison changes both model capacity and pretrained representations and therefore does not isolate the contribution of spiking computation. We construct ANN-VE as a continuous-valued counterpart of SVE, matching its architecture, parameter count, ImageNet pretraining data, and training budget, while keeping downstream fusion and action decoding unchanged. We additionally evaluate ANN-VE with softmax normalization applied to its attention scores.

VariantVELEParams (M)Avg. ↑
B0DINOv3-BBERT216.196.3
B1DINOv3-BSIE216.196.1
B2SVEBERT148.692.6
B3SVESIE148.692.4
B4ANN-VEBERT148.692.9
B5ANN-VESIE148.692.7
B6ANN-VE w/ softmaxBERT148.694.5
B7ANN-VE w/ softmaxSIE148.694.2

ANN-VE and SVE achieve comparable performance. Applying softmax to ANN-VE improves average success from 92.9% to 94.5% with BERT and from 92.7% to 94.2% with SIE, yielding gains of 1.6 and 1.5 %, respectively. These results suggest that the current SVE remains a key performance bottleneck, with its hardware-friendly, softmax-free SSA potentially limiting visual representation quality. Improving its attention mechanism is therefore a promising direction for further performance gains.

Weakness 4 and Question 4: Matched Evaluation and Consistent Success-Rate Aggregation. The Meta-World comparison uses different evaluation sample sizes, while LIBERO-Plus results use different aggregation conventions. Both issues require consistent protocols and explicit reporting.

Response: Thank you for highlighting the evaluation-budget mismatch. We re-evaluated Spike-driven VLA using 10 episodes per task, following SmolVLA’s initialization configuration and evaluation protocol. Under this matched budget, our method achieves 75.64% average success, exceeding SmolVLA (2.25B) by 7.40 % (75.64% vs. 68.24%).

Episodes/taskMethodSpike-drivenParams (B) ↓Easy ↑Medium ↑Hard ↑V. Hard ↑Avg. ↑
10TinyVLA×0.4277.621.511.415.831.6
10π₀×3.2071.848.241.730.047.9
10SmolVLA (0.45B)×0.4582.5041.8045.0060.0057.30
10SmolVLA (2.25B)×2.2587.1451.8270.0064.0068.24
10Spike-driven VLA (Ours)✓0.1583.2166.3685.0068.0075.64

For LIBERO-Plus, we have revised Tables 3 and 10 using a consistent aggregation protocol and evaluated our method using three different random seeds. The individual seed results, together with the mean ± sample standard deviation across seeds, are reported below. All statistics are computed without intermediate rounding and reported to one decimal place, yielding an overall mean success rate of 55.2%.

Perturbation categorySeed 7Seed 123Seed 48Mean ± sample SD
Camera Viewpoints41.240.240.940.8 ± 0.5
Robot Initial States46.749.448.148.1 ± 1.3
Language Instructions48.749.348.949.0 ± 0.3
Light Conditions84.284.484.084.2 ± 0.2
Background Textures61.361.861.461.5 ± 0.3
Sensor Noise34.535.835.235.2 ± 0.7
Object Layout67.567.767.467.5 ± 0.1
Overall (macro-average)54.955.555.155.2 ± 0.3

Reference:

[1] Xie H, Yao C, Wu X, et al. TurboVLA: Real-time vision-language-action model at 32 Hz on an RTX 4090 with< 1 GB VRAM[J]. arXiv preprint arXiv:2607.27205, 2026.

Reviewer_ErYaOpenReview ↗

Response

The rebuttal addresses most of my previous concerns. However, two points remain worth clarifying:​

  1. The matched ANN-VE baseline achieves slightly higher success rates than SVE. The authors should further clarify the computational efficiency trade-off between these two models to better justify the benefits of spiking computation.​
  2. The updated Meta-World evaluation uses 10 episodes per task, while Appendix E still reports 50 trials per task. This inconsistency should be corrected in the revised manuscript.

Thank you for your positive assessment and constructive suggestions, which are valuable for improving our manuscript. We will expand the appendix to clarify the computational efficiency trade-offs between SVE and ANN-VE.

For a 224×224 input, the estimated theoretical energy of ANN-VE and the reported energy reference for the original SVE backbone are summarized below.

ModelParameters (M)Theoretical Energy (mJ/image) ↓
ANN-VE≈19.035.88
SVE (original backbone reference)≈19.05.90

As shown above, this reference comparison suggests approximately 83.6% lower arithmetic energy for the spiking backbone, highlighting the potential energy benefits of sparse spiking computation at a comparable parameter budget.

We will also promptly correct the oversight in the appendix to clarify that the Meta-World evaluation uses 10 episodes per task.

Thank you once again for the time and effort you have devoted to reviewing our manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.