Pith. sign in

REVIEW 3 major objections 6 minor 34 cited by

This paper claims that asynchronous vision-language-action robot control becomes accurate and stable when the policy is fed the robot state at the moment its new actions will actually start executing, a future state computed by rolling the

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 19:17 UTC pith:J77D7TEB

load-bearing objection A clean, promising async-inference system whose named mechanism—future-state conditioning—is not actually isolated from the co-introduced offset fine-tuning, so the paper's causal story is under-supported even though the engineering is solid. the 3 major comments →

arxiv 2512.01031 v2 pith:J77D7TEB submitted 2025-11-30 cs.RO cs.AIcs.LG

VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference

Jiaming Tang , Yufei Sun , Yilong Zhao , Shang Yang , Yujun Lin , Zhuoyang Zhang , James Hou , Yao Lu
show 2 more authors
Zhijian Liu Song Han
This is my paper
classification cs.RO cs.AIcs.LG
keywords asynchronous inferencevision-language-actionaction chunkingfuture statetemporal offsetaction quantizationrobot controllatency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Vision-language-action models (VLAs) that plan robot actions are usually run synchronously: the robot freezes while the model thinks, causing stalls and slow reactions. Running the model in the background while the robot moves fixes latency but creates a mismatch: by the time the model's new action chunk is ready, the robot is in a different state than the one the model saw. VLASH argues this mismatch can be closed without extra computation or architecture changes by feeding the model the future robot state at the start of the next execution interval, computed by rolling the current state forward with the actions still being executed. The paper shows this future-state conditioning, aided by a temporal-offset fine-tuning trick and optional action quantization, preserves or improves accuracy while cutting reaction latency up to 17.4x and task time up to 2.03x. If true, it would let large VLAs handle dynamic, physically interactive tasks that synchronous pipelines cannot.

Core claim

The central claim is that under asynchronous inference, the VLA's prediction-execution gap is primarily a robot-state gap, not just a sensory gap: while the model computes, the robot keeps executing the previous action chunk, so the state at execution start is fully determined by the current state plus those pending actions. VLASH computes that future state s_{t+Δ} = s_t + sum of the remaining actions and conditions the policy on the pair (observation at inference start, future state) instead of (observation, current state). Because large VLAs under-use state input, the paper adds a temporal-offset fine-tuning scheme that trains the model on pairs (o_t, s_{t+δ}, a_{t+δ...}) with a fixed obse

What carries the argument

The load-bearing identity is s_{t+Δ} = s_t + sum_{i=t}^{t+Δ-1} a_i: the robot state at the start of the execution interval equals the current state plus the actions from the previous chunk that will still be running during the inference delay. This rollforward is computed with no extra model call and fed to the policy as the conditioning state. The second piece is temporal-offset augmentation: during fine-tuning, the same observation is paired with states and actions shifted by δ steps, and attention is masked so each offset branch attends only to the shared observation and its own tokens, making the model learn to use the future state while reusing the observation encoding (3.26x faster per

Load-bearing premise

The method assumes the robot state during the inference delay evolves by exact additive application of the previously issued actions, with no dynamics, slipping, contact forces, tracking error, or safety-layer modification changing that trajectory.

What would settle it

A concrete experiment: run the same VLASH deployment on a robot with known low-level tracking error—for example, a compliant arm or one with friction—and measure the distance between the computed future state s_{t+Δ} and the measured state at execution start. If that distance exceeds roughly the magnitude of a single action step, the claimed accuracy improvement should reverse. Alternatively, ablate by using a future state computed from random actions or from measured executed actions; if accuracy against naive async collapses when the rollforward is inaccurate, the load-bearing premise is con

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Asynchronous inference can be made accurate without architectural changes or per-step runtime overhead, so VLAs can be deployed on dynamic, fast-reaction tasks that were previously infeasible.
  • Reaction latency drops to roughly the model inference time, up to 17.4x faster than synchronous inference, because the robot never waits for the model.
  • Quantizing groups of fine-grained micro-actions into coarser macro-actions yields 1.5-2.0x task speedup with minimal accuracy loss, with the trade-off tunable by the quantization factor.
  • Temporal-offset fine-tuning makes the model robust across different inference delays and preserves synchronous performance, so a single deployment works on hardware with different compute budgets.
  • Large VLAs can achieve physically interactive behaviors like ping-pong rallies, which the paper demonstrates for the first time.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The rollforward identity is deterministic and additive; on robots with low-level tracking error, slip, contact forces, or safety-layer modifications, the computed future state will drift from the true state. A natural extension would feed a state estimate from a forward model or use measured executed actions instead of assuming exact kinematics.
  • Because the method forces the model to rely on state input, the same offset trick could compensate for other sources of delay, such as network latency in cloud robotics, by conditioning on a predicted state after transmission.
  • The paper keeps observations fixed during offset training and never predicts future visual input; extending the idea to forecast observations, or to handle occlusions, remains an open direction that the paper does not address.
  • The 17.4x reaction-latency figure assumes model inference is the only latency source; on systems where execution time is short relative to inference, the end-to-end speedup will be smaller and the accuracy-versus-speed trade-off will shift.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes VLASH, a framework for asynchronous inference of Vision-Language-Action models. The central idea is to condition the policy on a rolled-forward future robot state s_{t+Δ}, computed as the current state plus the remaining actions from the previously issued action chunk, after fine-tuning with temporal offsets on states and actions. This is intended to bridge the prediction–execution gap caused by inference delay. The paper also contributes an efficient shared-observation fine-tuning scheme and an action-quantization method for additional speed. Experiments on Kinetix and LIBERO report accuracy improvements over naive asynchronous inference (up to 30.5% on Kinetix), and real-world experiments on π0.5 report up to 2.03× task speedup and large reaction-latency reductions, including dynamic tasks such as ping-pong and whack-a-mole.

Significance. If the mechanism is validated, VLASH is practically significant: it offers a simple, low-overhead route to asynchronous VLA control and enables dynamic tasks that are infeasible with synchronous inference. The paper's strengths include a clean formulation, open-source code, real-hardware demonstrations, and a genuinely elegant shared-observation fine-tuning trick that gives a 3.26× training speedup. The core causal claim, however, is not yet isolated: the paper does not demonstrate that conditioning on the rolled-forward future state, rather than the temporal-offset fine-tuning alone, is responsible for the reported gains. The roll-forward identity is also unvalidated in settings with contact, slip, or low-level tracking error. Therefore, the significance is conditional on additional ablations and roll-forward-error quantification.

major comments (3)
  1. [§4.2, Table 1, Fig. 6] The central mechanism is not isolated. Section 4.2 states that π0.5 'under-utilizes the robot state' and that fine-tuning without state input outperforms fine-tuning with state input on LIBERO, and that simply feeding a future state at test time is insufficient. Yet the reported improvements over naive async (e.g., 30.5% at Δ=4 on Kinetix) are consistent with an alternative explanation: temporal-offset fine-tuning alone—training the model to predict action chunks shifted by Δ from a stale observation—could produce the gains, with the state channel contributing little. A minimal ablation set is needed: (a) offset fine-tuning with the current state at inference, (b) offset fine-tuning with the rolled-forward future state (VLASH), and (c) standard fine-tuning with future-state conditioning. Table 1 and Fig. 6 do not include these variants, so the paper's title-level claim that future-state
  2. [§4.1, Fig. 3(c)] The roll-forward identity s_{t+Δ}=s_t+Σ a_i assumes that executed actions are exact additive state deltas with no dynamics, slip, contact forces, low-level tracking error, or safety-layer modification during the inference interval. In contact-rich manipulation and dynamic tasks, the true state at execution start can differ materially from this open-loop prediction; conditioning on an inaccurate future state could degrade rather than improve stability. The manuscript reports no comparison between the rolled-forward state and the actual execution-start state, and no experiment comparing VLASH with the true future state (available in simulation) against the rolled-forward state. Such an experiment would quantify sensitivity to roll-forward error and is necessary to establish the mechanism.
  3. [§4.4, Fig. 7] The speedup claim 'up to 2.03× with preserved accuracy' depends on the quantization factor q being chosen after observing task outcomes. The text says 'we select task-dependent quantization factors that maintain success rates close to the unquantized policy,' but no protocol is given for selecting q without access to test rollouts. Since q is a free parameter tuned to the reported data, the speed-accuracy trade-off is demonstrated but the 'preserved accuracy' speedup is not a predictive claim. Please report a fixed q per task chosen before evaluation, a validation-based selection rule, or a sweep with confidence intervals over held-out rollouts.
minor comments (6)
  1. [Abstract] The abstract as reproduced at the top of the manuscript reports 'up to 11.8×' reaction-latency reduction and '1.5-2.0× task completion speedup,' while the full-text abstract reports 'up to 17.4×' and '2.03×.' Please harmonize these numbers.
  2. [Abstract / §7.4] The main text claims 'no architectural changes,' but Appendix §7.4 describes an optional state-projection layer for π0.5 that 'consistently improves control smoothness.' Please clarify whether the reported real-world results use this modification and soften the 'no architectural changes' claim accordingly.
  3. [Fig. 6] Kinetix success rates are reported without error bars or confidence intervals, despite 1,024 rollouts per data point. Please include standard errors or confidence intervals to support comparisons between VLASH, RTC, and naive async.
  4. [§5.1.1] Kinetix experiments use a 4-layer MLP-Mixer policy rather than a VLA. This is reasonable as a general test of the method, but the paper should be explicit that this result does not directly evaluate a vision-language-action model.
  5. [§5.2.2, Table 2] The reaction-latency comparison defines asynchronous latency as inference latency only. This is a worst-case assumption that should be stated more carefully: it presumes the event is captured at the start of an inference cycle and that the new action chunk can begin immediately upon completion of inference, with no additional scheduling delay.
  6. [General] There are minor typos, e.g., 'evalute' in §5.1.2 and 'additonal' in §2. Please proofread the manuscript.

Circularity Check

0 steps flagged

No significant circularity: the core accuracy claims are held-out empirical results rather than by-construction identities.

full rationale

The derivation chain is not circular. The roll-forward state s_{t+Δ}=s_t+Σa_{t:t+Δ-1} (Sec. 4.1, Fig. 3c) is an input computation from the previously issued action chunk; it is not fitted to, or derived from, the target action chunk being predicted. The temporal-offset fine-tuning (Sec. 4.2) trains on ground-truth future chunks from trajectories, and the reported success rates are measured on 1,024 Kinetix rollouts per data point and 16 real-world rollouts per method, so the improvements are not enforced by construction. In fact, the paper's own Table 1 shows that fine-tuning without state input outperforms fine-tuning with state input on LIBERO (Sync w/o state 97.7% vs Sync 96.8%), demonstrating that state conditioning is not trivially beneficial; whether the rolled-forward future state is actually used is an empirical question, not a tautology. The one self-citation (AWQ [22] in the action-quantization analogy) is not load-bearing for any central claim. The skeptical concern that gains might come from offset augmentation alone, or that the additive roll-forward identity is inaccurate under contact-rich dynamics, is a correctness/identifiability limitation of the experiments, not a circularity: the paper does not define the predicted accuracy in terms of its training labels, its own definitions, or a self-citation chain. Latency speedups are arithmetic consequences of the asynchronous inference definitions, but the paper reports them as measurement outcomes and does not dress them up as derived predictions of the model. Overall, the central claims are self-contained against external benchmarks and do not reduce to their inputs.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central method rests on two domain assumptions that are structural rather than standard: deterministic additive state roll-forward and the learnability of future-state conditioning. The quantization factor is a fitted parameter chosen after observing task success, while the delay range and deployment delay are hardware-driven choices. No new physical entities are introduced.

free parameters (3)
  • Training offset range Δmax = 3 (LIBERO fine-tuning; Kinetix delays up to 4)
    Chosen by hand as the maximum inference delay the model is trained to handle. Central to the offset augmentation but not fitted to success rates.
  • Deployment inference delay Δ = 4 steps (RTX 4090 laptop); 1.52-3.21 steps across GPUs
    Hardware-dependent delay used to compute the roll-forward state at test time. Must match the training offset range.
  • Action quantization factor q = 2 and 3, selected per task
    Section 4.4 states 'we select task-dependent quantization factors that maintain success rates close to the unquantized policy', then uses those chosen q values to report speedups. This is a post-hoc selection, inflating the apparent speed-accuracy tradeoff.
axioms (5)
  • domain assumption The robot state transition is deterministic and additive: s_{t+Δ}=s_t + a_t + ... + a_{t+Δ-1} for delta-position actions
    Used in Sec 4.1, Fig 3c (s3 = s1 + a1 + a2). Fails for velocity/torque actions, contact, slip, external forces, or low-level tracking error.
  • domain assumption The previously issued action chunk will be executed exactly as generated during the inference interval [t, t+Δ)
    Required for the roll-forward computation; any safety layer, chunk switching, or execution modification breaks the predicted future state.
  • ad hoc to paper Temporal-offset fine-tuning makes the VLA actually use the future state input and generalize across delays
    The paper asserts in Sec 4.2 that simply feeding a future state to a standard model is insufficient, motivating the offset augmentation, but provides no direct ablation isolating this mechanism.
  • domain assumption The current observation o_t remains a sufficient visual input when paired with the future state, despite being stale by Δ
    The paper acknowledges the future environment is unknown (Sec 4.1) but relies on the model compensating via body state, mirroring human reaction delay.
  • domain assumption Action quantization does not violate robot actuation limits: a macro-action summing q micro-actions is executable within one control tick
    Sec 4.4 defines macro-actions as sums of fine-grained deltas; this assumes velocity/position limits are not binding, which may fail at high q.

pith-pipeline@v1.3.0-alltime-deepseek · 17201 in / 14409 out tokens · 142963 ms · 2026-08-03T19:17:43.732664+00:00 · methodology

0 comments
read the original abstract

Vision-Language-Action models (VLAs) are becoming increasingly capable across diverse robotic tasks. However, these models are typically deployed under synchronous inference, where the robot waits for model inference to complete before acting, and cannot perceive or respond to environmental changes during action execution. This not only introduces noticeable action stalls, but also significantly increases reaction latency, fundamentally limiting the applicability of VLAs to dynamic, real-time tasks. Asynchronous inference offers a promising solution to achieve continuous and low-latency control by enabling robots to execute actions and perform inference simultaneously. However, because the robot and environment continue to evolve during inference, a temporal misalignment arises between the prediction and execution intervals. This leads to significant action instability, while existing asynchronous methods either degrade accuracy or introduce runtime overhead to mitigate it. We propose VLASH, a simple yet effective method for asynchronous VLA inference that delivers smooth, accurate, and fast reaction control without architectural changes or additional runtime overhead. VLASH leverages the future execution-time state by rolling the robot state forward with the previous action chunk, thereby bridging the gap between prediction and execution. Experiments show that VLASH reduces reaction latency by up to 11.8x compared to synchronous inference and consistently outperforms all asynchronous baselines in accuracy. With action quantization, it further achieves 1.5-2.0x task completion speedup with minimal accuracy loss. Moreover, it empowers state-of-the-art VLAs such as $\pi_{0.5}$ to handle fast-reaction, high-precision tasks including playing ping-pong and playing whack-a-mole, where traditional synchronous inference fails. Code is available at https://github.com/mit-han-lab/vlash

Figures

Figures reproduced from arXiv: 2512.01031 by James Hou, Jiaming Tang, Shang Yang, Song Han, Yao Lu, Yilong Zhao, Yufei Sun, Yujun Lin, Zhijian Liu, Zhuoyang Zhang.

Figure 1
Figure 1. Figure 1: VLASH enables VLA to play ping-pong rallies with humans. Snapshots showing 𝜋0.5 [16] with VLASH successfully tracking and striking a fast-moving ping-pong ball during a rally. The robot initiates its reaction by the third frame, demonstrating low-latency perception-to-action response. The task requires both fast reaction and smooth continuous motion, which are enabled by our asynchronous inference with fut… view at source ↗
Figure 2
Figure 2. Figure 2: Prediction-execution misalignment in asynchronous inference. Due to inference delay Δ, the model predicts actions for the prediction interval [𝑡, 𝑡 + 𝐾) but they execute during the execution interval [𝑡 + Δ, 𝑡 + Δ + 𝐾). inference for the next one. Because the execution duration of an action chunk is typically longer than the model inference time, the robot can immediately switch to the next chunk once the … view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between VLASH and existing methods. (a) Synchronous inference: the robot stalls during inference, introducing slow reactions. (b) Naive async: the model predicts based on stale state 𝑠1 while execution begins at future state 𝑠3, causing misalignment and discontinuity. (c) VLASH rolls forward the robot state (𝑠3 = 𝑠1 + 𝑎1 + 𝑎2) and condition on the execution-time state, achieving fast reaction an… view at source ↗
Figure 5
Figure 5. Figure 5: Action quantization for efficient execution. We group consecutive fine-grained micro-actions into coarser macro-actions to accelerate robot motion. The original trajectory with fine￾grained actions 𝑎0, 𝑎1, 𝑎2, . . . (gray) is quantized into a shorter trajectory with macro-actions ˆ𝑎0, 𝑎ˆ1, 𝑎ˆ2, 𝑎ˆ3 (black), where each macro-action summarizes 𝑞 consecutive fine-grained actions (e.g., 𝑎ˆ0 = 𝑎0 + 𝑎1 + 𝑎2 for … view at source ↗
Figure 6
Figure 6. Figure 6: Performance on Kinetix benchmark. We evaluate the success rate under different execution horizons 𝐾 and inference delays Δ. Left: Fixed inference delay Δ = 1 with varying execution horizon 𝐾. Right: Execution horizon adapts to inference delay, i.e., 𝐾 = max(Δ, 1), with varying Δ. For the Sync baseline, inference delay is always Δ = 0, but the execution horizon 𝐾 follows the same settings as other baselines… view at source ↗
Figure 7
Figure 7. Figure 7: Real-world evaluation results on manipulation tasks. We evaluate 𝜋0.5 [16] on three tasks with different inference methods. Left: Score percentages (based on 2-point scoring: 1 for success of picking up the object, 1 for task completion) of VLASH and baselines across three tasks. Right: Task completion times with green arrows indicating speedup of VLASH (𝑞=2) relative to synchronous baseline. VLASH (𝑞) app… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 34 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. The Speedup Paradox: Rethinking Inference Speed-Quality Trade-off in Embodied Tasks

    cs.RO 2026-06 unverdicted novelty 7.0

    TISED framework reveals paradoxical effects where inference optimizations can lengthen task completion time on static tasks or raise success rates on dynamic tasks in embodied AI.

  2. Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection

    cs.RO 2026-06 unverdicted novelty 7.0

    PAINT reframes asynchronous flow-based action chunking as an initial noise selection problem solved via backward Euler inversion and a repainting rule.

  3. DSSP: Diffusion State Space Policy with Full-History Encoding

    cs.RO 2026-05 conditional novelty 7.0

    DSSP is a history-conditioned diffusion state space policy that uses SSMs to encode full observation streams with an auxiliary dynamics objective and hierarchical fusion, achieving SOTA results with reduced model size...

  4. See What Matters: Differentiable Grid Sample Pruning for Generalizable Vision-Language-Action Model

    cs.RO 2026-05 conditional novelty 7.0

    GridS is a plug-and-play differentiable module for geometry-aware visual token resampling in VLA models that achieves under 10% token retention and 76% FLOPs reduction with no success-rate loss.

  5. DiscreteRTC: Discrete Diffusion Policies are Natural Asynchronous Executors

    cs.RO 2026-04 unverdicted novelty 7.0

    Discrete diffusion policies support native asynchronous execution via unmasking for real-time chunking, delivering higher success rates and 0.7x inference cost versus flow-matching RTC on dynamic robotics benchmarks a...

  6. DiscreteRTC: Discrete Diffusion Policies are Natural Asynchronous Executors

    cs.RO 2026-04 unverdicted novelty 7.0

    Discrete diffusion policies act as natural asynchronous executors for robotics by treating action generation as iterative unmasking, yielding higher success rates and lower computation than flow-matching real-time chu...

  7. Action Chunk Scheduling for Batched Robot Policy Serving

    cs.RO 2026-07 conditional novelty 6.0

    A lookahead scheduler that simulates each robot's action-queue state before choosing batches improves throughput in heterogeneous multi-robot policy serving by up to 18% in real-world tests.

  8. FBFM: A Training-Free Asynchronous Feedback Mechanism for Flow-Matching in World-Action Models Execution

    cs.RO 2026-07 conditional novelty 6.0

    A training-free masked pseudoinverse-guidance mechanism injects newly observed states and committed actions into the active flow-matching chunk of world-action models, improving long-horizon execution on selected mani...

  9. $\pi\mathbf{R}^2$: Reactive Real-time Flow Policies

    cs.RO 2026-07 conditional novelty 6.0

    πR² makes flow-matching VLA policies reactive by splitting conditioning into fresh proprioception and stale vision-language features and using a one-step-per-call staircase noise schedule, reaching ~25 Hz closed-loop ...

  10. FutureRTC: Real-Time Robot Execution with Anticipatory-Conditioned Action Chunking

    cs.RO 2026-07 conditional novelty 6.0

    A frozen VLA policy can stay accurate under asynchronous execution by consuming adapter-predicted execution-time observation latents and states instead of stale ones.

  11. Token-Wise Latent Streaming from Slow Reasoners to Fast Planners for Dynamic Vision Language Navigation

    cs.RO 2026-07 conditional novelty 6.0

    Streaming intermediate hidden states from a slow VLM to a fast flow-matching planner, token by token, improves dynamic social VLN success and reduces observation staleness.

  12. Reflex: Real-Time VLA Control through Streaming Inference

    cs.RO 2026-07 conditional novelty 6.0

    Reflex caches timestep-invariant perception features in flow-matching VLA models to deliver ~2.58x inference speedup and stable 50Hz streaming control.

  13. ROSA: A Robotics Foundation Model Serving System for Robot Factories

    cs.RO 2026-07 unverdicted novelty 6.0

    ROSA introduces shared GPU-pool serving, robotics-aware abstractions for multi-model pipelines, and factory-productivity scheduling that improves output by up to 12.06x over dedicated per-robot systems.

  14. Communication-Aware Robot Execution for Cloud Inference under Spatially Heterogeneous Connectivity

    cs.RO 2026-06 unverdicted novelty 6.0

    A communication-aware execution method uses a request-response window and available connectivity maps to select request points during primitive execution, yielding best or tied-best task success with fewer attempts an...

  15. The Speedup Paradox: Rethinking Inference Speed-Quality Trade-off in Embodied Tasks

    cs.RO 2026-06 unverdicted novelty 6.0

    TISED decomposes inference optimization effects on embodied tasks and identifies paradoxical outcomes where faster per-step inference can increase task completion time on static tasks or raise success rates on dynamic tasks.

  16. DREAM-Chunk: Reactive Action Chunking with Latent World Model

    cs.RO 2026-06 unverdicted novelty 6.0

    DREAM-Chunk uses test-time sampling and latent-world-model rollouts to select robust action chunks from chunking-based VLA policies, improving performance under stochastic dynamics on simulation and hardware tasks.

  17. Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack

    cs.RO 2026-06 conditional novelty 6.0

    A full learning stack—10K-hour UMI data, a flow-matching VLA, preference-optimization RL, and asynchronous deployment—reports SOTA RoboTwin results and cross-embodiment transfer to four real robots.

  18. DAM-VLA: Decoupled Asynchronous Multimodal Vision Language Action model

    cs.RO 2026-06 unverdicted novelty 6.0

    DAM-VLA decouples per-modality temporal processing in vision-language-action models via latent buffers refreshed at sensor rates, achieving 95.2% average success versus 40.95% for synchronous baselines on seven real-w...

  19. Learning to Assist: Collaborative VLAs for Implicit Human-Robot Collaboration

    cs.RO 2026-06 conditional novelty 6.0

    VLA models with inference-time steering mitigate action leakage in implicit human-robot collaboration, supporting longer horizons and yielding faster, more reliable assembly than shorter-horizon baselines in a 16-pers...

  20. Intercepting the Future: Latent-Space Predictive World Model for Dynamic VLA Manipulation

    cs.RO 2026-06 unverdicted novelty 6.0

    AHEAD augments frozen VLAs with a 4.9M-parameter latent world model that forecasts future visual features using optical-flow motion cues, achieving 79-97% success on dynamic simulation tasks and high real-robot succes...

  21. DEFLECT: Delay-Robust Execution via Flow-matching Likelihood-Estimated Counterfactual Tuning for VLA Policies

    cs.RO 2026-05 unverdicted novelty 6.0

    DEFLECT is an offline post-training method that improves async VLA policy success rates under high inference delays by using flow-matching likelihood ratios on counterfactual fresh/stale action pairs from a frozen ref...

  22. See What Matters: Differentiable Grid Sample Pruning for Generalizable Vision-Language-Action Model

    cs.RO 2026-05 unverdicted novelty 6.0

    GridS reduces visual tokens in VLA models to under 10% of the original count via task-aware differentiable resampling, delivering 76% lower FLOPs with no drop in task success rate on benchmarks and real robots.

  23. AsyncShield: A Plug-and-Play Edge Adapter for Asynchronous Cloud-based VLA Navigation

    cs.RO 2026-04 unverdicted novelty 6.0

    AsyncShield restores VLA geometric intent from latency via kinematic pose mapping and uses PPO-Lagrangian to balance tracking with LiDAR safety constraints in a plug-and-play module.

  24. FASTER: Rethinking Real-Time Flow VLAs

    cs.RO 2026-03 unverdicted novelty 6.0

    FASTER adds a Horizon-Aware Schedule to flow VLAs that compresses immediate-action denoising to one step while keeping long-horizon trajectory quality, lowering real-robot reaction latency.

  25. FASTER: Rethinking Real-Time Flow VLAs

    cs.RO 2026-03 conditional novelty 6.0

    FASTER uses a horizon-aware flow sampling schedule to compress immediate-action denoising to one step, slashing effective reaction latency in real-robot VLA deployments.

  26. TIDAL: Temporally Interleaved Diffusion and Action Loop for High-Frequency VLA Control

    cs.RO 2026-01 conditional novelty 6.0

    TIDAL raises VLA control feedback from ~2.4 Hz to ~9 Hz by caching semantic intent and interleaving one-step flow generation with execution, doubling dynamic interception success in simulation.

  27. Action ControlNet: A Lightweight Delay-Aware Adapter for Smooth Asynchronous Control in Vision-Language-Action Models

    cs.RO 2026-06 unverdicted novelty 5.0

    Action ControlNet is a lightweight delay-aware adapter that conditions mostly frozen VLA action heads on executed motion suffixes to reduce discontinuities in asynchronous robot trajectories.

  28. Slow Brain, Fast Planner: Latency-Resilient VLM-Augmented Urban Navigation

    cs.RO 2026-06 unverdicted novelty 5.0

    A training-free fusion layer enables stale VLM selections to improve a real-time planner's trajectory scoring for urban sidewalk navigation, yielding 30% ADE reduction in challenging scenarios.

  29. LAGO Policy: Latency-Aware Asynchronous Diffusion Policies with Goal-Directed Collision-Free Planning for Smooth Manipulation

    cs.RO 2026-06 unverdicted novelty 5.0

    LAGO Policy proposes a unified framework combining latency-aware classifier-free guidance, goal prediction from demonstrations, and spatial-temporal trajectory optimization to enable smooth collision-free execution in...

  30. Real-Time Execution with Autoregressive Policies

    cs.RO 2026-06 unverdicted novelty 5.0

    Autoregressive VLA policies achieve real-time execution via tokenization horizon adjustment and constrained decoding, outperforming flow-matching policies in speed and performance across simulated and real environments.

  31. DexPIE: Stable Dexterous Policy Improvement from Real-World Experience

    cs.RO 2026-06 unverdicted novelty 5.0

    DexPIE improves dexterous manipulation success rates by 37% over demo policies via real-world experience collection with adapted intervention, multi-stage DAgger, asynchronous relative-action inference, and optimality...

  32. Understanding Asynchronous Inference Methods for Vision-Language-Action Models

    cs.RO 2026-05 unverdicted novelty 5.0

    Controlled benchmarks show per-step residual correction (A2C2) as most effective for VLA asynchronous inference up to d=8 delays on Kinetix with over 90% solve rate, outperforming inpainting and conditioning while tra...

  33. Causal World Modeling for Robot Control

    cs.CV 2026-01 unverdicted novelty 5.0

    LingBot-VA combines video world modeling with policy learning via Mixture-of-Transformers, closed-loop rollouts, and asynchronous inference to improve robot manipulation in simulation and real settings.

  34. On-Device Robotic Planning: Eliminating Inference Redundancy for Efficient Decision-Making

    cs.RO 2026-05 unverdicted novelty 4.0

    REIS reduces inference redundancy in embodied robotic planning via lightweight gating and routing while preserving task performance on ALFRED and real robots.

Reference graph

Works this paper leans on

44 extracted references · 14 linked inside Pith · cited by 30 Pith papers

  1. [1]

    Gemini robotics 1.5: Push- ing the frontier of generalist robots with advanced embod- ied reasoning, thinking, and motion transfer.arXiv preprint arXiv:2510.03342, 2025

    Abbas Abdolmaleki, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Arenas, Ash- win Balakrishna, Nathan Batchelor, Alex Bewley, Jeff Bing- ham, Michael Bloesch, et al. Gemini robotics 1.5: Push- ing the frontier of generalist robots with advanced embod- ied reasoning, thinking, and motion transfer.arXiv preprint arXiv:2510.03...

    Pith/arXiv arXiv 2025

  2. [2]

    Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Chourdia, Will Constable, Alban Desmaison, Zachary De- Vito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshiteej Kalam- barkar, Laurent Kirsch, Micha...

    2024

  3. [3]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.𝜋 0: A vision-language- action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024. 2, 5

    Pith/arXiv arXiv 2024

  4. [4]

    Real- time execution of action chunking flow policies.arXiv preprint arXiv:2506.07339, 2025

    Kevin Black, Manuel Y Galliker, and Sergey Levine. Real- time execution of action chunking flow policies.arXiv preprint arXiv:2506.07339, 2025. 1, 2, 3, 6, 8

    Pith/arXiv arXiv 2025

  5. [5]

    Lerobot: State-of- the-art machine learning for real-world robotics in pytorch

    Remi Cadene, Simon Alibert, Alexander Soare, Quentin Gallouedec, Adil Zouitine, Steven Palma, Pepijn Kooij- mans, Michel Aractingi, Mustafa Shukor, Dana Aubakirova, Martino Russi, Francesco Capuano, Caroline Pascal, Jade Choghari, Jess Moss, and Thomas Wolf. Lerobot: State-of- the-art machine learning for real-world robotics in pytorch. https : / / github...

  6. [6]

    Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Ab- hishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, Albert Tung, Alex Bew- ley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anchit Gupta, Andrew Wang, Andrey Kolobov, Anikait Singh, Animesh Garg, Aniru...

    2025

  7. [7]

    Flashattention-2: Faster attention with better paral- lelism and work partitioning, 2023

    Tri Dao. Flashattention-2: Faster attention with better paral- lelism and work partitioning, 2023

    2023

  8. [8]

    Fu, Stefano Ermon, Atri Rudra, and Christopher R ´e

    Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher R ´e. Flashattention: Fast and memory-efficient exact attention with io-awareness, 2022

    2022

  9. [9]

    Flex attention: A programming model for generating optimized attention kernels, 2024

    Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex attention: A programming model for generating optimized attention kernels, 2024

    2024

  10. [10]

    Ren, Homer Walke, Quan Vuong, Lucy Xiaoyang Shi, and Sergey Levine

    Danny Driess, Jost Tobias Springenberg, Brian Ichter, Lili Yu, Adrian Li-Bell, Karl Pertsch, Allen Z. Ren, Homer Walke, Quan Vuong, Lucy Xiaoyang Shi, and Sergey Levine. Knowl- edge insulating vision-language-action models: Train fast, run fast, generalize better, 2025. 6

    2025

  11. [11]

    GPTQ: Accurate post-training compres- sion for generative pretrained transformers.arXiv preprint arXiv:2210.17323, 2022

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. GPTQ: Accurate post-training compres- sion for generative pretrained transformers.arXiv preprint arXiv:2210.17323, 2022. 5

    Pith/arXiv arXiv 2022

  12. [12]

    Galaxea AI Technology Co

    Ltd. Galaxea AI Technology Co. Galaxea ai.https:// galaxea-ai.com/cn, 2025. 7

    2025

  13. [13]

    Galaxea AI Technology Co

    Ltd. Galaxea AI Technology Co. R1 lite.https:// galaxea- ai.com/cn/products/R1- Lite, 2025. 7

    2025

  14. [14]

    Effi- cient vision-language-action models for embodied manipula- tion: A systematic survey.arXiv preprint arXiv:2510.17111, 2025

    Weifan Guan, Qinghao Hu, Aosheng Li, and Jian Cheng. Effi- cient vision-language-action models for embodied manipula- tion: A systematic survey.arXiv preprint arXiv:2510.17111, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Hugging Face

    Inc. Hugging Face. So-101.https://huggingface. co/docs/lerobot/en/so101, 2025. 7

    2025

  16. [16]

    1, 2, 3, 5, 6, 7, 8

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al.𝜋 0.5: a vision-language-action model with open-world generaliza- tion.arXiv preprint arXiv:2504.16054, 2025. 1, 2, 3, 5, 6, 7, 8

    Pith/arXiv arXiv 2025

  17. [17]

    The better you learn, the smarter you prune: Towards efficient vision-language-action models via differentiable to- ken pruning, 2025

    Titong Jiang, Xuefeng Jiang, Yuan Ma, Xin Wen, Bailin Li, Kun Zhan, Peng Jia, Yahui Liu, Sheng Sun, and Xianpeng Lang. The better you learn, the smarter you prune: Towards efficient vision-language-action models via differentiable to- ken pruning, 2025

    2025

  18. [18]

    Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Bal- akrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, Peter David Fagan, Joey Hejna, Masha Itk- ina, Marion Lepert, Yecheng Jason Ma, Patrick Tree Miller, Jimmy Wu, Suneel Belkhale, Shivin Dass, Huy Ha, Arhan Jain, Abraham Lee,...

    2025

  19. [19]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Fos- ter, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024. 2

    Pith/arXiv arXiv 2024

  20. [20]

    Fine-tuning vision-language-action models: Optimizing speed and suc- cess, 2025

    Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and suc- cess, 2025

    2025

  21. [21]

    Pytorch distributed: Expe- riences on accelerating data parallel training.arXiv preprint arXiv:2006.15704, 2020

    Shen Li, Yanli Zhao, Rohan Varma, Omkar Salpekar, Pieter Noordhuis, Teng Li, Adam Paszke, Jeff Smith, Brian Vaughan, Pritam Damania, et al. Pytorch distributed: Expe- riences on accelerating data parallel training.arXiv preprint arXiv:2006.15704, 2020. 8

    Pith/arXiv arXiv 2006

  22. [22]

    Awq: Activation-aware weight quantization for llm compression and acceleration

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for llm compression and acceleration. InMLSys,

  23. [23]

    Libero: Benchmarking knowledge transfer for lifelong robot learning.arXiv preprint arXiv:2306.03310, 2023

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning.arXiv preprint arXiv:2306.03310, 2023. 4, 6, 1 10

    Pith/arXiv arXiv 2023

  24. [24]

    Running vlas at real-time speed.arXiv preprint arXiv:2510.26742, 2025

    Yunchao Ma, Yizhuang Zhou, Yunhuan Yang, Tiancai Wang, and Haoqiang Fan. Running vlas at real-time speed.arXiv preprint arXiv:2510.26742, 2025. 1, 2

    arXiv 2025

  25. [25]

    Kinetix: Investigating the training of general agents through open-ended physics-based control tasks

    Michael Matthews, Michael Beukman, Chris Lu, and Jakob Foerster. Kinetix: Investigating the training of general agents through open-ended physics-based control tasks. 2025. 2, 6

    2025

  26. [26]

    GR00T N1: An open foundation model for generalist humanoid robots

    NVIDIA, Nikita Cherniadev Johan Bjorck andFernando Casta˜neda, Xingye Da, Runyu Ding, Linxi ”Jim” Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, Joel Jang, Zhenyu Jiang, Jan Kautz, Kaushil Kundalia, Lawrence Lao, Zhiqi Li, Zongyu Lin, Kevin Lin, Guilin Liu, Edith Llon- top, Loic Magne, Ajay Mandlekar, Avnish Narayan, Soroush Nasiriany, Scott Reed, ...

    2025

  27. [27]

    Fast: Efficient action tokenization for vision- language-action models, 2025

    Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision- language-action models, 2025

    2025

  28. [28]

    Spatialvla: Exploring spatial repre- sentations for visual-language-action model, 2025

    Delin Qu, Haoming Song, Qizhi Chen, Yuanqi Yao, Xinyi Ye, Yan Ding, Zhigang Wang, JiaYuan Gu, Bin Zhao, Dong Wang, and Xuelong Li. Spatialvla: Exploring spatial repre- sentations for visual-language-action model, 2025

    2025

  29. [29]

    Leave no observation behind: Real-time correction for vla action chunks.arXiv preprint arXiv:2509.23224, 2025

    Kohei Sendai, Maxime Alvarez, Tatsuya Matsushima, Yutaka Matsuo, and Yusuke Iwasawa. Leave no observation behind: Real-time correction for vla action chunks.arXiv preprint arXiv:2509.23224, 2025. 1, 2, 3

    arXiv 2025

  30. [30]

    Flashattention-3: Fast and accurate attention with asynchrony and low-precision, 2024

    Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, and Tri Dao. Flashattention-3: Fast and accurate attention with asynchrony and low-precision, 2024

    2024

  31. [31]

    Smolvla: A vision-language-action model for afford- able and efficient robotics.arXiv preprint arXiv:2506.01844,

    Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Ar- actingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. Smolvla: A vision-language-action model for afford- able and efficient robotics.arXiv preprint arXiv:2506.01844,

    Pith/arXiv arXiv

  32. [32]

    Hitter: A humanoid table tennis robot via hierarchical plan- ning and learning.arXiv preprint arXiv:2508.21043, 2025

    Zhi Su, Bike Zhang, Nima Rahmanian, Yuman Gao, Qiayuan Liao, Caitlin Regan, Koushil Sreenath, and S Shankar Sastry. Hitter: A humanoid table tennis robot via hierarchical plan- ning and learning.arXiv preprint arXiv:2508.21043, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Galaxea g0: Open-world dataset and dual- system vla model.arXiv preprint arXiv:2509.00576v1, 2025

    Galaxea Team. Galaxea g0: Open-world dataset and dual- system vla model.arXiv preprint arXiv:2509.00576v1, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Gemini robotics: Bringing ai into the physical world.arXiv preprint arXiv:2503.20020,

    Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Arenas, Travis Armstrong, Ashwin Balakrishna, Robert Baruch, Maria Bauza, Michiel Blokzijl, et al. Gemini robotics: Bringing ai into the physical world.arXiv preprint arXiv:2503.20020,

    Pith/arXiv arXiv

  35. [35]

    Mlp- mixer: An all-mlp architecture for vision.Advances in neural information processing systems, 34:24261–24272, 2021

    Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lu- cas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp- mixer: An all-mlp architecture for vision.Advances in neural information processing systems, 34:24261–24272, 2021. 6

    2021

  36. [36]

    Tinyvla: Towards fast, data-efficient vision-language-action models for robotic ma- nipulation, 2025

    Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, and Jian Tang. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic ma- nipulation, 2025

    2025

  37. [37]

    Smoothquant: Accurate and ef- ficient post-training quantization for large language models,

    Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and ef- ficient post-training quantization for large language models,

  38. [38]

    A survey on efficient vision-language-action models, 2025

    Zhaoshu Yu, Bo Wang, Pengpeng Zeng, Haonan Zhang, Ji Zhang, Lianli Gao, Jingkuan Song, Nicu Sebe, and Heng Tao Shen. A survey on efficient vision-language-action models, 2025

    2025

  39. [39]

    Dreamvla: A vision-language-action model dreamed with comprehensive world knowledge, 2025

    Wenyao Zhang, Hongsi Liu, Zekun Qi, Yunnan Wang, Xin- qiang Yu, Jiazhao Zhang, Runpei Dong, Jiawei He, Fan Lu, He Wang, Zhizheng Zhang, Li Yi, Wenjun Zeng, and Xin Jin. Dreamvla: A vision-language-action model dreamed with comprehensive world knowledge, 2025

    2025

  40. [40]

    Cot-vla: Visual chain-of-thought reasoning for vision-language-action mod- els, 2025

    Qingqing Zhao, Yao Lu, Moo Jin Kim, Zipeng Fu, Zhuoyang Zhang, Yecheng Wu, Zhaoshuo Li, Qianli Ma, Song Han, Chelsea Finn, Ankur Handa, Ming-Yu Liu, Donglai Xi- ang, Gordon Wetzstein, and Tsung-Yi Lin. Cot-vla: Visual chain-of-thought reasoning for vision-language-action mod- els, 2025

    2025

  41. [41]

    Learning fine-grained bimanual manipulation with low- cost hardware.arXiv preprint arXiv:2304.13705, 2023

    Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low- cost hardware.arXiv preprint arXiv:2304.13705, 2023. 1

    Pith/arXiv arXiv 2023

  42. [42]

    Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn

    Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low- cost hardware, 2023. 3

    2023

  43. [43]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control

    Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. InConference on Robot Learning, pages 2165–2183. PMLR, 2023. 2 11 VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference Supplementary...

    2023

  44. [44]

    Appendix 7.1. SmolVLA Results on LIBERO Benchmarks To further evaluate the generalization of VLASH across different VLAs, we conduct additional experiments on SmolVLA-450M [31], a compact yet efficient vision- language-action model. Following the same experimental setup as described in Sec. 5.1.2, we fine-tune SmolVLA on the LIBERO benchmark [23] for 30K ...