Efficient VLA Inference

VLA: Vision-Language-Action Model

Overview

This topic studies how to run vision-language-action models fast enough for responsive, reliable robot control, including on edge devices. Common approaches prune or cache visual tokens, quantize or distill models, reduce action-generation steps, and overlap inference with observation and execution. Methods are tested on manipulation benchmarks such as LIBERO and RoboTwin, as well as real robots, using measures like latency, control frequency, memory, energy, and task success. The main tension is that faster inference can undermine closed-loop reliability or motion quality, so gains must be assessed in full task execution rather than by model speed alone.

Latest papers

221 papers36 papers in the last four weeks, down 35% on the four weeks before. 0.3% of all new papers.

Oct 26, 2026

ASPEN: Adaptive Semantic Reasoning via Command Persistence Prediction for Efficient Navigation VLAs

Navigation vision-language-action (VLA) policies repeatedly invoke costly semantic reasoning even when the next command remains unchanged. Command reuse can reduce computation but risks missing instruction-dependent turns or stops. We introduce ASPEN, an event-triggered scheduler based on command persistence prediction: estimating whether the frozen base policy would reissue its previous command. A lightweight multimodal predictor combines temporal observations, language instructions, execution history, and cached VLA confidence. At eligible command boundaries, sufficient predicted persistence permits command reuse. Otherwise, ASPEN triggers fresh semantic reasoning from the VLA. Evaluated with frozen NaVILA on cross-dataset RxR-CE val-unseen, ASPEN reduces VLA calls by 9.5% and estimated inference time by 8.1%, including scheduling overhead. It maintains comparable navigation performance and preserves identical baseline action sequences on 92.1% of episodes.
Oct 26, 2026

VLA-HWBench: Where the Time Goes When VLA Models Run on Edge Hardware

Vision-language-action (VLA) models increasingly run on edge devices aboard robots, where inference latency constrains closed-loop robot control. However, it remains unclear how to design VLA models, their inference software and edge hardware so that robots can run VLA inference in real time. Most VLA models report only task success, while prior studies cover limited action-generation strategies, divide inference into at most three stages, and lack comparisons of inference optimizations on the same models. Therefore, we introduce VLA-HWBench, a benchmark that measures 18 open VLA models on NVIDIA Jetson AGX Thor and AMD Ryzen AI Max+ PRO 395, with a seven-stage inference breakdown. Our study reveals three insights. First, action-generation strategy matters more than model size. It dominates 14 of 19 Jetson configurations, latency varies by 82×82\times on Jetson and 157×157\times on AMD, and parameter count explains at most 15% of latency variance. Models that generate long action chunks in a few flow-matching steps keep up with control, whereas token-by-token decoding does not. Second, kernel fusion with 4-bit weights helps most, running π0.5π_{0.5} and GR00T N1.7 2.6-2.9×\times faster than PyTorch with torch.compile, but it exists for only these two models and only on the Jetson. Optimization also moves the time to the stages it shortens least, as CPU preprocessing grows from 12% to 39% of the GR00T N1.7 call. Third, after optimization, memory bandwidth limits only the stages that process few tokens on the Jetson, such as the denoising steps, whereas software limits all other stages, from unfused kernels to serial preprocessing on the host CPU. Overall, low-latency edge VLAs should use few action-generation iterations, inference software should extend kernel fusion and low-bit weights to more models and platforms, and edge hardware needs high memory bandwidth and a fast host CPU.
Oct 25, 2026

TRIAGE: Efficient Vision-Language-Action Inference Needs Both Reuse and Pruning

Vision-language-action (VLA) models are increasingly used for robot manipulation because they map camera images and a language instruction directly to actions. Robots that cannot be tethered to a workstation, such as mobile manipulators and humanoids, must run them on onboard embedded boards, where the prefill latency of the language backbone grows with the hundreds of visual tokens processed at every control step. To meet real-time limits, training-free accelerators recompute only a subset of the visual tokens and skip the rest, either by pruning them or by reusing their cached key-value (KV) states. The two options save nearly the same computation, yet existing methods apply only one of them to every skipped token. Through a token-level diagnostic, we show that which option keeps the action closer to full recomputation depends on the token's drift, the change in its shallow-layer value since its cache was computed. On both backbones tested, reusing tokens below a drift threshold was no worse than pruning them, whereas beyond it, reuse yields up to 4.4x the action error of pruning; the same diagnostic calibrates this threshold per backbone. Building on this finding, we propose TRIAGE, which deems a cache valid when its drift is below this threshold and it is not too old. TRIAGE recomputes a fixed budget of tokens, prioritizing important tokens with invalid caches, reuses every other token with a valid cache, and prunes the rest; prune-only and reuse-only methods are the two extremes of this partition. Because the budget fixes the number of recomputed tokens, it directly sets the prefill latency. On LIBERO and SimplerEnv, TRIAGE stays within 3 pp of full recomputation in average success rate while cutting prefill latency on a Jetson AGX Orin by up to 48%, and with only 7% of visual tokens recomputed, it beats the strongest baseline by 24.4 pp on LIBERO-Object.
Oct 24, 2026

Power-of-Two Leaky Integrate-and-Fire Neurons for Event-Driven Vision-Language-Action Inference

Vision-language-action (VLA) models have demonstrated strong capabilities in robotic manipulation, yet their reliance on dense, continuous-valued computation presents a fundamental challenge for energy-constrained deployment. Inspired by the sparse, event-driven signaling and temporal integration mechanisms of biological neurons, a promising alternative is to introduce spiking computation into VLA models. However, enabling such computation without restructuring pretrained policies or compromising closed-loop control remains an open challenge. To address this challenge, we introduce the Power-of-Two Leaky Integrate-and-Fire (PoT-LIF) neuron, which emits signed spikes with power-of-two amplitudes. By combining magnitude-coded spike events with residual-preserving membrane dynamics, PoT-LIF enables continuous information to be progressively represented through discrete spikes. The Heartbeat formulation further localizes this process within weighted projections, bridging biologically inspired spiking computation and pretrained VLA architectures without modifying their external inference interfaces. We integrate PoT-LIF into the vision projection, vision-language projection, and action projection of π0.5π_{0.5} and evaluate both operating modes over 2,000 paired closed-loop LIBERO episodes per configuration. Standard-mode PoT-LIF VLA achieves success rates of 96.90% and 96.35% at T=1T = 1 and T=2T = 2, respectively, while Heartbeat mode achieves 97.25% and 96.90% at T=2T = 2 and T=4T = 4, compared with 96.70% for the original continuous-activation π0.5π_{0.5} baseline. At T=1T = 1 with a 60.01% firing rate, PoT-LIF uses approximately 1/7.67 of the dense-reference arithmetic energy, corresponding to a modeled reduction of 86.95%.
Oct 23, 2026

Coupling Execution Horizon and Sampling Depth in Flow-Matching VLA Policies

Flow-matching Vision-Language-Action (VLA) models generate expressive continuous trajectories for robotic manipulation, yet balancing real-time responsiveness with execution stability remains a persistent challenge. Standard deployment workflows typically treat numerical solver depth and execution horizon as decoupled choices. The horizon amortizes inference latency, while solver depth is kept high to preserve integration fidelity. In this paper, we reveal a fundamental compute-execution coupling in flow-matching VLAs governed by a systematic sign reversal. While coarse numerical integration degrades precision under rapid replanning, it markedly improves task performance under extended open-loop commitment. Geometrically, numerical truncation contracts lateral action dispersion around an invariant trajectory centroid, serving as an implicit regularizer that suppresses compounding open-loop drift. Across LIBERO, RoboTwin 2.0, and real-world Franka manipulation on representative VLA backbones including π0π_0, π0.5π_{0.5}, and GR00T, exploiting this joint coupling consistently achieves superior task success while drastically lowering inference latency, entirely training-free and without architectural modifications. These findings establish numerical sampling depth as an operational variance regulator, shifting its role from a static precision budget to an active deployment knob in robot learning. Project page: https://jiabingyang01.github.io/Horizon-Depth-VLA/
Oct 22, 2026

When One Step Is Enough: An Empirical Study of Euler Integration in Flow-Matching Policies

Flow matching is now the dominant approach for vision-language action (VLA) robot policies, thanks to its expressiveness and simple training objective. However, inference for standard flow matching requires iteratively solving an ODE, limiting the control frequency at deployment. While several works have proposed novel methods to allow for single-shot action generation, we ask whether such additional objectives are actually necessary. We find that across six flow matching VLAs, one-step action generation largely matches multi-step sampling at scale: multi-step integration confers little to no consistent advantage at inference. We empirically find that this holds as long as models are sufficiently trained and have enough capacity. Practically, the implication is that for large VLAs, the single-step policy comes essentially for free from the multi-step model, enabling low-latency deployment without retraining, distillation, or auxiliary objectives.
Oct 21, 2026

AGQ: Action-Guided Ultra-Low-Bit Quantization for Vision-Language-Action Models

Vision-language-action (VLA) models enable robots to perform tasks specified through natural-language instructions, but their large weight footprint challenges deployment on memory-constrained onboard platforms. Ultra-low-bit quantization reduces this footprint, yet preserving task performance and controlling weight recovery costs remain challenging. We propose Action-Guided Quantization, which combines action sensitivity with a measured proxy for weight recovery cost to allocate matrix bitwidths under storage budgets. Per-channel calibration compensates for quantization error, while a uniform bitwidth within each matrix preserves regular layouts. Bit-dependent grouped lossless coding reduces resident weight storage, and block-wise recovery limits temporary buffers. On LIBERO, AGQ achieves 92.7% task success at W3, exceeding AWQ and GPTQ by 1.2 and 3.7 percentage points, respectively, and retains 70.5% success at W2.5. An asymmetric configuration, denoted W1.5∗^{*}, uses W4 vision weights and a parameter-weighted average of 1.5 bits for language weights. With lossless coding, it achieves 6.10×6.10\times complete-model compression relative to BF16 while retaining 25.5% task success. Code and experimental results are available at https://github.com/Qiuxu00/AGQ
Oct 20, 2026

PACT-VLA: Speculative Action-Chunk Verification with Pre-Commit Observations

Vision-language-action (VLA) policies are commonly deployed with asynchronous inference in real-world settings, where the next action chunk is generated while the robot is still executing the previous one. By the time the new chunk is committed, the robot and scene have changed, so the chunk is conditioned on a stale observation and task success drops sharply as this inference delay grows. Generating multiple candidate chunks could let newer observations inform the final choice, but existing multi-candidate methods either rank candidates with the stale inference-start observation or repeatedly switch between candidates during execution, which can make the executed action sequence discontinuous. Our key insight is that inference delay can be used as decision time. While candidate generation is still underway, candidate selection can reuse the VLA itself to choose the complete action chunk that best fits a later observation. We propose PACT-VLA, a training-free framework that generates candidates from the inference-start observation and selects one using this pre-commit observation. Its scoring function, FlowRes, measures the residual between each candidate's implied flow direction and the velocity predicted by the frozen policy under the pre-commit observation, requiring no separate verifier. A pipelined implementation overlaps observation processing with candidate generation and batches scoring, so selection adds no control-step delay on the evaluated hardware. At a four-step delay, PACT-VLA improves average success by 15.0 percentage points over RTC on Kinetix and by 5.2 points over single-candidate asynchronous execution with GR00T N1.7 on LIBERO, and it further improves existing delay-robust methods. On three real-robot tasks, it raises average success from 60% to 82% over naive asynchronous execution with the same frozen π0.5π_{0.5} checkpoint.
Oct 19, 2026

SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models

The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate. Reducing these demands through aggressive compression, however, can degrade manipulation performance beyond what supervised fine-tuning alone recovers. We present SPAR (Structural Pruning and Action Recovery), which simultaneously prunes the hidden and feed-forward intermediate dimensions of the large language model (LLM) backbone through pruning-aware tuning, then applies reinforcement learning (RL) to recover policy performance. Applied to π0.5π_{0.5}, SPAR reduces the language backbone from 2.34B to 0.09B parameters and the full model from 3.37B to 1.12B, removing approximately 96% and 67% of their parameters, respectively. At this compression level, RL raises average success across the four LIBERO suites from 84.55% to 93.45%, an 8.90-percentage-point gain. The compressed policy achieves a 1.98×1.98\times full-policy inference speedup over the unpruned model. These results identify substantial intermediate-dimension redundancy and show that post-pruning RL can recover manipulation performance in compact policies.
Oct 19, 2026

PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models

Fast and reliable manipulation is crucial for practical robot deployment. Although Vision-Language-Action (VLA) models have recently shown remarkable progress as generalist robot policies, their execution often remains temporally inefficient. A key source of this inefficiency is that policies learn fine-grained temporal patterns from teleoperated demonstrations, requiring many low-level action steps. Temporal action quantization offers a simple way to accelerate execution by aggregating consecutive low-level actions into temporally coarser actions. However, applying action quantization without considering when fine-grained control is required introduces a speed-accuracy trade-off: accelerated actions may efficiently reduce redundant steps in some phases, while discarding critical motion details in precision-sensitive interactions. To address this, we propose PaceVLA, a unified framework for adaptive temporal execution in VLAs that integrates two coupled components: (1) mode-specific action decoders built on a shared action expert, supporting multiple temporal granularities and prediction horizons, and (2) a quantizability-guided mode selection scheme which decides whether coarse execution is appropriate, and a context-conditioned router that selects an execution mode at each inference step. Experiments across three multi-task simulation benchmarks and six real-world tasks show that PaceVLA improves execution efficiency while maintaining task success rate across multiple embodiments, including long-horizon and dexterous manipulation.
Oct 18, 2026

VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse

Flow-matching vision-language-action (VLA) models generate an action chunk by iteratively denoising Gaussian noise, but typically execute only a short prefix before replanning with updated observations. Under this receding-horizon paradigm, the unexecuted remainder---the action tail---is discarded, and the next chunk is regenerated from fresh noise. We show that this discarded tail retains transferable denoising progress: action-tail initialization can match or outperform a native 10-step Gaussian baseline using fewer denoising steps. However, the optimal denoising budget varies substantially across samples, indicating that tail reuse should be adapted to the updated environment. Based on this observation, we introduce VLA-Rolling, a lightweight inference framework that reuses action tails through three stages: action-tail inheritance, per-action confidence estimation, and confidence-guided refinement through renoising and denoising budget allocation, while the underlying VLA remains frozen. Across π0.5π_{0.5}, SmolVLA, X-VLA, and MolmoAct2 on LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0, VLA-Rolling reduces action-chunk inference cost in all 16 model--benchmark pairs, achieving up to a 1.48×1.48\times speedup, while improving success rate in 14 pairs by up to 15.0815.08 percentage points. These results show that action tails are not merely discarded predictions, but reusable intermediate states that enable more efficient and accurate VLA inference. These results demonstrate that the action tail can support the generation of better actions with fewer denoising steps. Our project is available at https://chenbao123ww.github.io/VLA-Rolling/.
Oct 17, 2026

Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing

Vision-Language-Action (VLA) models offer a promising paradigm for autonomous driving. However, the massive parameter scale of VLAs strains onboard computational resources during real-time inference. This directly increases energy consumption and limits vehicle battery range, making it critical to minimize computational overhead without compromising driving capability and explainability. Current research on efficient VLA-based driving focuses on sequential adaptation (e.g., Chain-of-Thought and token adjustment), leaving structural adaptation underexplored despite parameter redundancy across dynamic driving scenes. To address this gap, we propose Scene-Adaptive VLA, an efficient framework leveraging dynamic layer routing to adjust active parameters based on real-time scenes. Inspired by the temporal continuity of scenes and model predictive uncertainty, our approach characterizes model-perceived scene complexity via frame-level temporal scene variations and decision deviations. To uncover latent scene-to-parameter correspondences, we encode this characterization into scene-aware tokens, alongside two specialized queries. A budget-conditioned layer routing mechanism evaluates these refined queries to allocate a budget that determines the active parameter ratio and subsequently selects appropriate layer combinations. The entire framework is optimized end-to-end. Extensive evaluations on the Bench2Drive benchmark show that when adapted to a state-of-the-art VLA model, our approach significantly reduces inference-time computational overhead by 53.5% with negligible degradation in driving and language capabilities. Code is available at https:github.com/yang-luvu/Scene-Adaptive-VLA.
Oct 17, 2026

FlashRT-PAI: Fast Kernels Are Not Fast Pipelines for Physical AI

Vision-language-action (VLA), video-action (VA), and world action models (WAMs) form the backbone of modern physical AI systems, yet their real-time single-batch inference remains a defining bottleneck. Latency is dominated not by individual kernel computation, but by system-level overhead: thousands of short kernel launches, intermediate data movement across operator boundaries, and redundant recomputation across consecutive decisions. General-purpose graph compilers cannot address these costs, as they optimize individual calls in isolation and lack visibility into cross-call state semantics and global pipeline structure. We present FLASHRT-PAI, a low-latency inference engine built around model-hardware pipeline co-design rather than isolated kernel optimization. FLASHRT-PAI combines three core mechanisms: deep kernel fusion with launch overlap eliminates inter-operator boundaries and launch gaps; an execution contract formalizes the pipeline's stage DAG, update rates, and named buffers to enable cross-decision state reuse and selective refresh; and execution-state capsules enable fast snapshot, restore, and fork of persistent pipeline state. The system also supports drop-in acceleration of unmodified PyTorch models via parity- and latency-gated kernel replacement. Evaluation covers six policies on RTX 5090 and four on Jetson AGX Thor. At matched 16-bit precision, FLASHRT-PAI is 1.3-3.7x faster than official compiled paths. Combining pipeline co-design with reduced precision delivers 3.38-12.66x speedup over official eager implementations. In LIBERO-10, four-bit pi0.5 achieves 94.04% success on Thor and 93.32% on RTX 5090, compared with 92.4% for the measured official FP16 reference. FLASHRT-PAI is open source, distributed via Hugging Face Kernel Hub, and its opt-in pi0.5 kernels have been adopted into the LeRobot framework.
Oct 16, 2026

MambaAct: Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models

Vision-language-action (VLA) models have shown strong promise for generalist robotic manipulation, but their inference cost remains a major bottleneck for real-time control. This limitation is especially pronounced in flow-matching VLAs such as π0.5π_{0.5}, where the action expert is evaluated repeatedly during iterative denoising. We introduce MambaAct, an efficient Mamba action expert for flow-matching VLA policies. MambaAct retains the pretrained visual-language prefix encoder of π0.5π_{0.5}, while replacing its transformer-based action expert with a compact Mamba-based action expert. Within this expert, Mamba serves as the action mixer that models dependencies across the noisy continuous-action sequence. To preserve visual-language conditioning without repeated token-level cross-attention, MambaAct uses a lightweight Context Compressor to map Vision-language model (VLM) prefix tokens into a compact conditioning signal that gates and modulates each Mamba action block. MambaAct uses a 10×10\times smaller action expert and reduces flow-matching denoising latency by 64.8%64.8\% on LIBERO and 58.8%58.8\% in real-world deployment, while improving mean LIBERO success by 1.081.08 percentage points across three independent fine-tuning seeds and maintaining gains on LIBERO-Plus and real-world benchmarks. Our code and model weights are publicly available at https://github.com/giakhuyendihoc/MambaAct.
Oct 15, 2026

FastVLA: Task-Conditioned and Temporally Adaptive Visual Token Compression for Efficient Vision-Language-Action Models

Vision-Language-Action (VLA) models present a promising paradigm for general-purpose robot manipulation, but closed-loop deployment is often limited by inference efficiency: at each timestep, high-resolution, often multi-view images are encoded into hundreds of visual tokens that dominate transformer computation and latency. Token compression via attention is a natural solution but is challenging for VLA control because (1) decision-critical tokens are determined by the task, not generic visual saliency, and (2) the token budget varies substantially across steps, so static metrics and budgets waste computation or discard control-critical tokens. We introduce FastVLA, a training-free, task-aware, and temporally adaptive visual token compression framework. At each step, FastVLA scores token importance via task-to-vision attention and adaptively sets the compression rate from the importance-distribution uniformity, measured by an inverse-Simpson metric with online sliding-window calibration. We evaluate FastVLA on OpenVLA, OpenVLA-OFT, and CogACT across LIBERO and SIMPLER, and on a physical Franka Research 3 robot. In simulation, it reduces LLM-backbone FLOPs by 35-59% and action latency by 9-32% (a 1.09×\times-1.46×\times speedup) while maintaining task success, achieving a comparable or better efficiency-performance trade-off than existing plug-and-play token compression/pruning/reuse baselines. On the real robot, it cuts action-prediction latency by 20% (1.25×\times speedup) at comparable success. Code is available at https://github.com/jiuyixu25/FastVLA.
Oct 15, 2026

Fast-VLA: You Only Think Once for Robot Manipulation

Vision-language-action (VLA) policies commonly place a multi-billion-parameter model inside the feedback loop, repeatedly recomputing task context whenever they predict a new action chunk. Yet task understanding changes far more slowly than the robot state. We introduce Fast-VLA, which invokes a frozen 4B VLM once per episode to contextualize the instruction and initial views as fixed, multi-layer key-value task memory. During execution, a lightweight online actor fuses learned time- and arm-indexed queries with current visual and proprioceptive feedback to form action queries. These queries retrieve task context from the fixed memory and predict action chunks without re-running the foundation model. Fast-VLA achieves 92.46% and 92.40% success on RoboTwin 2.0 Clean and Randomized, and 95.95% and 79.11% on LIBERO and LIBERO-Plus evaluation. Against three continually conditioned baselines on the same hardware, it reduces RoboTwin amortized policy inference time by 10.9-17.1x and requires 1.22 GiB of post-prefill online resident memory. Task-memory swaps further show that the actor relies on episode-matched task memory. These results show that task context can be established once while closed-loop control remains responsive throughout execution. Project page: https://garfield-kh.github.io/FastVLA
Oct 15, 2026

LTM: Latent Transition Based In-Context Memory for Vision-Language-Action Models

Robotic manipulation often requires remembering interactions that are no longer visible. We introduce Latent Transition Memory (LTM), which stores history as compact latent transitions between consecutive observations. Instantiated on π0.5π_{0.5} as LTM-ππ, our method uses a frozen latent action encoder, a lightweight adapter, and causal KV caching for incremental inference without memory-specific semantic annotations or external reasoning-model calls. With DreamDojo transitions, LTM-ππ achieves 58.54% success on RoboMME and 80.5% on LIBERO-Mem, performing best among baselines without external reasoning models or memory-specific annotations. Ablations support the contribution of transition content beyond temporal cues and the efficiency of delayed memory admission. Scaling RoboMME demonstrations tenfold improves success from 57.67% to 87.50%, while a Cup Swap diagnostic highlights the role of process supervision in temporal composition. Under a 200-token memory budget, success remains 84.00%, versus 87.50% without compression. Finally, real-robot counting success improves from 11.25% to 82.50%. Together, these results demonstrate the effectiveness, scalability, and efficiency of latent transition memory, highlighting its capacity for complex temporal information, bounded-context operation, and real-world robotic control.
Oct 15, 2026

When to Ask for Help: A Decision-Theoretic Account and Measurement Standard for Slow-Module Scheduling in Dual-System Vision-Language-Action Models

Dual-system vision-language-action (VLA) models pair a high-frequency fast policy with a slow, expensive module, and published trigger mechanisms for invoking it lack a shared objective, measured costs, and any standard separating schedule quality from module quality. We provide the missing account: invocation is formalized as a calibrated decision problem whose fixed-ratio, fixed-budget, and history-gated schedules are degenerate cases of one quantity, the Value of Intervention (VoI); a renewal baseline proves the optimal schedule periodic under state-independent collapse, with adaptive gains governed by the cost ratio ρρ and the dispersion of collapse. Measured systems span three orders of magnitude in ρρ (FiS-VLA 0.950.95; RoboICL about 10310^3), and a direct test confirms the low-ρρ prediction: calibrated gains vanish under homogeneous collapse and grow with heterogeneity. Fourteen closed-loop arms on LIBERO test the contested middle of the continuum (ρ=1.53ρ=1.53): the adaptive trigger allocates its queries exactly as the threshold rule predicts, yet the language-rewrite channel cannot convert them into success on either suite, and a revert actuator that undoes a failed rewrite is the only intervention with zero net downside on both LIBERO-Spatial and LIBERO-Goal, where the rewrite deficit deepens from thirteen to forty-two points. A four-measure leak-free protocol cuts apparent early-warning AUC from 0.930.93 to 0.620.62 on 929929 deployment traces, a detection-limited frontier theorem converts detection quality into a schedulability statement, and a compilation loop turns measured value functions into per-task refresh periods within 2.4%2.4\% of the closed-loop oracle. Together these results turn slow-module invocation from a design opinion into a measured decision.
Oct 14, 2026

Causal Analysis of Robot Foundation Models Enables Efficiency Gains

Vision-language-action (VLA) models are a promising approach to general purpose robot manipulation. While the internal computation of vision-language models, which compose the backbone of VLAs, has been studied extensively, less is known about how their perceptual and semantic representations are transformed into action in VLA models. We study this question mechanistically in three frontier VLA models (π0.5π_{0.5}, MolmoAct2, and MolmoBot) by analyzing the causal pathways through which the input image and the instruction shape actions. Across the three models, activation patching shows that causally relevant information influences the action expert (AE) through a narrow band of mid-to-late cross-attention layers, which we term the action-conditioning interface. The content of this information changes with token and layer within the model. Early-to-mid layer representations of the target noun token determine which object is selected, whereas mid-to-late layers carry the spatial information in text tokens that redirects the movement when patched. Where spatial information is stored in prompt tokens various based on the architecture. In MolmoAct2 the model forms task-conditioned representations at the text positions that follow the instruction, whereas π0.5π_{0.5} additionally stores this information at image token positions. We show that these findings can be translated directly into improving model efficiency. We find that restricting MolmoBot's action expert conditioning to only 3 out of 36 layers reduces inference FLOPs by up to 33%33\% while maintaining comparable task success with the default model across 3,000 evaluation episodes (46.7%46.7\% vs. 46.9%46.9\%), and the same principle carries over to real-robot deployment of π0.5π_{0.5}.
Oct 14, 2026

Accelerating VLA Inference with Perturbation Informed Execution

Action-chunking vision-language-action (VLA) models predict sequences of future robot actions. Under fixed-horizon execution, the robot executes a fixedlength prefix of each predicted chunk before invoking the policy again, leaving the remaining predictions unused. For a given number of executed actions, short execution horizons require more policy calls. Extending the execution horizon amortizes policy inference cost over more actions. However, uniformly applying long horizons across episodes can limit opportunities for precise corrections. To address this limitation, we propose PICES (Perturbation-Informed Chunk and Execution Horizon Selection), a training-free method that jointly selects an action chunk and the number of actions executed before the next policy call. Within a single policy call, the action expert generates a batch of action chunks from multiple initial noise samples and their locally perturbed counterparts, using the same conditioning features across all samples. PICES exploits differences between corresponding actions in each nominal chunk and its counterpart to determine candidate-specific execution horizons and select a nominal chunk for execution. Experiments on simulation benchmarks and real-world manipulation tasks demonstrate that PICES achieves speedup by up to 2.3x in wall-clock policy inference time on Jetson Orin and Thor relative to fixed horizon baselines while improving the success rate.
Oct 13, 2026

TacitVLA: Reasoning Vision-Language-Action Models Without Reasoning Latency

Textual reasoning before acting has improved vision-language-action (VLA) models, but at the cost of latency: every action conditioned on the reasoning text must wait for autoregressive decoding. We show that actions do not need to wait for or condition on the generated reasoning to benefit from it, because the reasoning supervision already improves the backbone representation. We introduce TacitVLA, a reasoning VLA whose action expert never attends to the reasoning it generates for the current call. On long-horizon tasks where the reasoning text also serves as episodic memory, it can generate each entry during action execution rather than before acting. We evaluate it in a controlled study on four benchmarks spanning manipulation and autonomous driving, memory and non-memory tasks: RoboMME, RMBench, LIBERO, and PhysicalAI-AV. On all four, TacitVLA matches the performance of its explicit-reasoning counterpart at up to 3.4x lower latency, and matches the latency of the same architecture trained without reasoning at up to +17.1% success. In absolute terms, it surpasses the best previously published results on the RoboMME Counting suite (82.7% vs. 66.8%) and RMBench (77.5% vs. 56.5%), and reaches 72.9% on LIBERO-Plus and a 52.5% scene score on the public split of the PhysicalAI-AV closed-loop driving challenge. Project page: https://tacitvla.github.io.
Oct 11, 2026

ΔΔ Flow Matching for Robot Foundation Models

Flow matching is a common generative recipe for robot foundation models. Vision-language-action models (VLAs) generate action chunks, while world-action models (WAMs) additionally generate future visual states. Both start generation from a fixed Gaussian prior, despite temporal continuity between past and future states and actions. This requires costly training and iterative sampling, especially for WAMs with an additional visual branch. We introduce ΔΔ-flow, which constructs the source from recent history with mild noise rather than from a fixed Gaussian prior. This moves the source closer to the target, reducing transport to a small residual ΔΔ that can be learned efficiently and enables single-step inference. We integrate ΔΔ-flow into the VLA policy π0.5π_{0.5} and the WAM policies ImageWAM and Fast-WAM. On LIBERO, ImageWAM reaches 90% success in one epoch rather than six, and Fast-WAM reaches 95% in two epochs rather than eight. With LoRA fine-tuning, π0.5π_{0.5} exceeds its baseline's five-epoch performance after one epoch. Single-step inference gives 1.54×1.54\times, 2.49×2.49\times, and 4.45×4.45\times speedups on π0.5π_{0.5}, ImageWAM, and Fast-WAM, respectively. It preserves action smoothness, achieving a 4.5×4.5\times reduction in action jerk compared with one-step Gaussian sampling. ΔΔ-flow reaches state-of-the-art generalization performance on LIBERO-Plus. Action-history sources can be applied to pretrained WAM policies without retraining. Replacing the Gaussian action source at inference largely preserves task success with a single sampling step and can slightly improve it. We further study how action representation and visual information in action learning affect history-to-future prediction in ΔΔ-flow. These results show that ΔΔ-flow is simple yet effective, improving the efficiency and generalization of robot foundation models.
Oct 11, 2026

Recovering Low-Bit Failures in Vision-Language-Action Policies via Quantization-Conditioned Reinforcement Learning

Quantization lowers deployment costs of vision-language-action (VLA) policies but can compromise task completion. Across five VLA models and 4,050 real-world trials, we systematically reveal that quantization-aware training (QAT) generally improves on post-training quantization (PTQ) but leaves substantial gaps to full precision (FP) on long-horizon tasks and under visual and semantic shifts, while its imitation-based objective does not explicitly use rollout success or partial completion. To exploit these outcome signals, we introduce Quantization-Conditioned Reinforcement Learning (QCRL), an offline RL method coupling progress-based advantage estimation with quantization-aware source conditioning. FP rollouts supply reference experience beyond action-level distillation, progress-scaled terminal penalty distinguishes partial completion, and quantization-aware conditioning selects advantageous reference behavior within the quantized policy. We evaluate weight-only INT4 W4A16 on π0.5π_{0.5} and GR00T N1.5 and joint weight-activation NVFP4 W4A4 on π0.5π_{0.5}. Using only in-distribution data, QCRL reaches 99.3% of FP's mean score across three visual-shift conditions for π0.5π_{0.5} with INT4 W4A16, alongside 42.9% model-memory savings. With NVFP4 W4A4, it reaches 100.7% of the same FP baseline, alongside a 1.73×1.73\times speedup using NVFP4/FP8 deployment. Hardware gains are relative to FP16 on Jetson Thor.
Oct 10, 2026

Barrier-Constrained Guidance For Real-time Asynchronous VLA Execution

While asynchronous execution enables VLA models to operate in real-time environments by predicting the next chunk during execution, it inevitably introduces boundary misalignment with the committed action stream. Existing delay-aware solutions address this at the prohibitive cost of extensive retraining, architectural modifications, or fixed-prefix inpainting that inflates computation when low latency is vital. To address these problems, we propose FreeBar, a training-free real-time inference plugin that enables robust asynchronous execution with only marginal additional inference cost by casting chunk alignment in flow-based VLAs as barrier-guided action-space constraint. Specifically, FreeBar introduces a lightweight Control Barrier Function (CBF)-inspired consistency constraint between consecutive action chunks and solves the resulting correction in closed form, avoiding costly backpropagation through the VLA action head. By progressively tightening the consistency bound during generation, FreeBar aligns generated chunks with the committed execution stream while preserving early-stage generative and within-chunk future action flexibility,. Across Kinetix, RoboCasa, and diverse real-world robotic platforms with multiple VLA backbones, FreeBar improves success rates over naive asynchronous execution by 10%--15% on average, with larger gains in highly dynamic scenarios. It also preserves real-time computational efficiency, decreasing inference latency per action chunk over 45% than RTC. Additionally, as a universal plugin, FreeBar remains robust across distinct latency environments, base policies, and hyperparameter settings.
Oct 9, 2026

Memory Without Curation: Learning Multi-Stream Latent Memory for Robot Policies

It is widely appreciated that robotic policies require long-term memory to perform complex tasks. However, directly conditioning policies on all past observations leaves policy learning susceptible to learning spurious shortcuts, and introduces large latency at policy deployment. As such, today's memory-augmented policies curate what to remember, either by retaining only recent frames, selected keyframes, or textual summaries, or by learning latent representations of the visual stream alone. In this work, we propose an alternative approach IMPRINT, a policy-agnostic latent memory framework that augments vision-language-action (VLA) models with multi-stream memory without explicitly constraining what the policy should remember. IMPRINT compresses different streams into a latent memory that is updated incrementally at inference, enabling low latency. To avoid shortcut learning, IMPRINT adopts a two-stage training procedure, first supervising memory to reconstruct multiple streams (observations, actions, proprioception) of past information, as well as estimate current task progress, followed by end-to-end finetuning, jointly with the policy, on action prediction. Across both simulation and real-world tasks demanding memory, IMPRINT improves average success over the base VLA by at least 26%, outperforming popular alternatives, and maintaining lowest inference latency among memory-augmented VLAs. Through careful ablations, we demonstrate the utility of incorporating diverse information streams into robotic memory, validating the advantages of foregoing history curation. Project website: https://yilin-wu98.github.io/imprint_website/.
Oct 8, 2026

REACT: Rolling Denoising and Dual Decoupling for Reactive Robot Control with VLA Models

Flow-based vision-language-action (VLA) models generate action chunks for temporally coherent robot motion, but chunked control creates a fundamental closed-loop trade-off: long chunks provide smooth execution, whereas frequent replanning improves reactivity at the cost of action discontinuities. We introduce REACT, a rolling-denoising framework that makes flow-based VLAs more reactive while preserving long-horizon context. Instead of regenerating entire action chunks from scratch, REACT maintains a persistent action buffer with staggered flow timesteps. At each control step, the full horizon is denoised using the latest observation, the cleanest action block is executed, partially refined future blocks are shifted forward, and fresh noise is appended to the tail. As a result, each executed action block is refined across multiple recent observations before deployment. To support real-time control, we further introduce dual decoupling, which separates sensing, VLM encoding, DiT denoising, and action execution, enabling high-frequency observation updates and action streaming under practical compute constraints. Across the RoboTwin 2.0 simulation benchmark and real-world tasks spanning bimanual manipulation and dynamic control on multiple robot platforms, REACT improves task success and reduces reaction latency while producing smoother trajectories than frequent-replanning and asynchronous baselines.
Oct 8, 2026

FastJEV: Understanding Redundancy for Compact JEV Inference

JEV models make multimodal decisions by directly scoring candidates. Although the common context is encoded once, candidate evaluation can still repeat matching token histories, duplicate inference states, and execute the full backbone. In this paper, we study these sources of redundancy and present FastJEV for compact candidate evaluation. We jointly organize history reuse and state storage, since sharing computation requires preserving states for later branches. We first introduce shared context anchoring to reuse recurrent initial states and omit unused final recurrent caches. We extend this reuse through candidate prefix sharing, retaining the intermediate states needed by subsequent branches. To further reduce the depth of these paths, we apply decision guided pruning based on relative score changes measured on a small unlabeled set. Our method retains full context encoding and all candidates without additional training. We evaluate FastJEV across three OmniJev model sizes on five public benchmarks and reconstructed LIBERO-10 offline questions. At the selected pruning budgets, the complete method reduces candidate depth by 43.75% to 45.83%, while retaining 93.66% to 97.52% of the original task scores on average across the six evaluation sets. Through controlled experiments, we show how candidate overlap and branching structure affect the execution cost of history reuse. In our implementation, candidate prefix sharing can reduce repeated computation while increasing latency. These findings motivate designing sharing granularity and execution schedules together for efficient JEV inference.
Oct 7, 2026

SpikingVLA: Asynchronous Spiking Vision-Language-Action Models

ANN-to-SNN conversion offers a practical route toward energy-efficient spiking Vision-Language-Action (VLA) models by bypassing the substantial cost of training large-scale SNNs from scratch. However, existing methods often require many timesteps to maintain competitive performance, resulting in substantial inference latency for real-time VLA deployment. To address this challenge, we introduce SpikingVLA, an ANN-to-SNN conversion framework that enables accurate and low-latency spiking VLA inference. Specifically, we propose a Dendritic Integrate-and-Fire (DIF) neuron that alleviates channel-wise activation outliers through dendritic mixing and adaptive somatic firing, enabling accurate ANN-to-SNN conversion with fewer timesteps. Building on DIF neurons, we further introduce an asynchronous execution mechanism that overlaps temporal computation across VLA components, reducing synchronization overhead and latency. Extensive experiments demonstrate that SpikingVLA achieves competitive navigation performance with substantially improved inference efficiency. Compared with existing spiking VLA methods, SpikingVLA improves SR and SPL by 11.9% and 12.6%, respectively, while reducing first-action latency by 11.2×\times. These results establish SpikingVLA as a practical framework for deploying pretrained VLA models with high-performance and low-latency spiking inference.
Oct 6, 2026

CARE: Certifying Acceleration for Vision-Language-Action Inference

While vision-language-action (VLA) models have advanced rapidly, running them at every control step remains expensive. Prior work accelerates VLA inference using techniques like action chunking and visual-token pruning, typically evaluating based on latency and average task success. However, acceleration may discard information and break tasks the original policy would solve, a risk hidden by average metrics. Measuring these failures is challenging because action deviations compound over closed-loop trajectories, meaning task failure is only observable across full episodes. We therefore define an acceleration-induced failure via paired rollouts from identical initial conditions, tracking when the reference succeeds but the accelerated policy fails. To manage this, we introduce CARE, an approach for certified accelerator selection. CARE uses paired rollouts on a calibration set to provide finite-sample guarantees that acceleration-induced failure risk stays below a user-specified budget. It deploys the fastest certified candidate, falling back to the reference if none qualify. By relying only on terminal outcomes and measured compute, CARE applies unchanged across diverse acceleration mechanisms, while sequential testing and failure-triggered reference rollouts keep certification affordable. On four LIBERO suites with OpenVLA-OFT, CARE certifies 9.09.0--10.8×10.8\times speedups while guaranteeing (at 95%95\% confidence) that at least 85.8%85.8\% of reference-solved episodes are preserved. Under tight budgets, selectors without guarantees exceed the budget in up to 75%75\% of trials, whereas CARE stays within budget and its sequential form uses 78.9%78.9\% fewer rollouts than exhaustive evaluation. CARE further generalizes to flow-step reduction for π0.5π_{0.5}, and to Qwen3.5-9B and Llama-3.1-8B agents in Crafter.
Oct 6, 2026

ActTune: Action-Aware Precision and GPU Operating-Point Adaptation for Energy-Efficient Vision-Language-Action Inference

Vision-language-action (VLA) policies repeatedly invoke inference to control robots, making graphics processing unit (GPU) energy a recurring cost of task execution. Reducing energy per inference call, however, may not reduce energy per successful task if numerical errors increase failures or slower inference prolongs execution. We therefore target GPU energy per successful task while preserving task success and keeping the inference-latency increase within 10%. Our approach builds on two observations: quantization sensitivity varies across action classes, model layers, and weights versus activations; and numerical precision changes the workload, shifting favorable GPU operating points. We introduce ActTune, an action-aware framework that connects layer-wise precision allocation with workload-dependent GPU operating-point selection over requested frequency--power-cap pairs. A lightweight decision tree learns its splits and leaf precision configurations directly from configuration action errors, then selects precision before each policy call. The controller forecasts the next workload and applies the selected GPU operating point asynchronously using a lookup table calibrated under a latency budget. A shared resident quantized weight bank enables configuration switching without weight reconstruction or additional policy evaluations. On LIBERO, a benchmark for lifelong robot learning, ActTune improves mean task success by up to 2.3% relative to state of the art. Relative to the original BF16 implementations, it delivers up to 2.02×2.02\times faster inference and, with GPU operating-point adaptation, reduces energy per successful task by up to 76.8%.