PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Houlong Xiong, Zhenqi Qiu, Zechen Wang, Suohang Zhang, Yiyu Ren, Wanting Xu, Hongfei Niu, Chengyang He, Ge Sun, Ran Cheng, Qian Zhu

arXiv:2610.12007v1cs.ROcs.AI

TL;DR

Chunked VLA control trades smooth, coherent motion for responsiveness to fresh observations, and existing approaches leave a gap between inference speed and continuous closed-loop control. REACT rolls denoising across a persistent action buffer and decouples sensing, encoding, denoising, and execution. Across simulation and real-world tasks, it improves success, reactivity, and trajectory smoothness over frequent-replanning and asynchronous baselines.

  • Problem

    Chunked VLA actions preserve temporal consistency but limit how quickly fresh observations can affect actions still inside a chunk, while latency reduction alone does not ensure fresh inputs and continuous executable outputs.

    Source: §1 · p. 2

  • Method

    REACT progressively refines a persistent action buffer under successive observations and uses dual-decoupled inference with matched staircase training.

    Source: §1 · p. 2

  • Results

    Across RoboTwin 2.0 simulation and real-world tasks, REACT improves success, reactivity, and smoothness; real-world success rises from 17.7% to 64.8% versus Async+RTC.

    Source: §1 · p. 2, §4.2 · p. 7

  • Takeaways & Limitations

    The results suggest real-time VLA control should be evaluated by how effectively it uses fresh observations, not only by model inference latency.

    Source: §5 · p. 9

    Validation uses only the flow-based VLA backbone π0.5, leaving cross-model compatibility empirically untested.

    Source: §6 · p. 9

Abstract

from arXiv · show

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.

1 Introduction

Chunked VLA control must balance smooth, coherent actions against responsiveness to fresh observations. REACT rolls denoising across observations, while dual decoupling and matched staircase training support deployment.

1 Introduction

  • Long action chunks preserve smooth, mode-consistent motion but reduce observation freshness, whereas frequent replanning improves reactivity but weakens temporal coherence.The introduction frames input and output throughput as system-level measures of closed-loop responsiveness beyond inference latency alone.Source: §1 · p. 2
  • REACT maintains a persistent action buffer and refines future blocks under successive observations instead of regenerating each chunk from scratch.Each executed block is refined across multiple recent observations while retaining a long action horizon.Source: §1 · p. 2
  • Across RoboTwin 2.0 simulation and real-world tasks, REACT improves the success–reactivity trade-off and trajectory smoothness over frequent-replanning and asynchronous baselines.The paper identifies dual-decoupled inference and matched staircase training as part of this approach.Source: §1 · p. 2

2 Related Work

Prior work addresses chunked VLA responsiveness through faster inference, asynchronous execution, or rolling generation, but these approaches leave challenges for fresh conditioning and continuity. REACT applies rolling denoising to VLA inference and separates input processing from action generation.

Chunked VLA policies and dynamic reactivity

  • Action chunking supports temporally consistent visuomotor control, but longer chunks refresh observations less often and can leave actions based on stale context after scene changes.Chunked policies produce action sequences rather than single commands and are used with models including diffusion and flow-matching policies.Source: §2 · pp. 2–3

Efficient inference acceleration and asynchronous inference

  • Inference acceleration and asynchronous execution reduce or hide latency, yet chunk switching can create observation-state mismatch, discontinuities, performance drops, and trajectory jitter.RTC-style methods condition future actions on committed or executed prefixes, while input throughput can remain tied to action-execution cadence.Source: §2 · p. 3

Rolling denoising for generative and robotic policies

  • Earlier rolling-denoising methods target vision, motion, or non-VLA robotic policies; REACT brings rolling buffers to VLA stacks and decouples input processing from action generation.The passage identifies coupled VLM encoding and DiT denoising as bottlenecks for VLA inference.Source: §2 · p. 3

3 Methodology

REACT rolls denoising across a persistent action buffer, decouples sensing and VLM encoding from denoising and execution, and trains on the same staircase flow-time schedule used at inference. Together, these mechanisms retain long-horizon context while incorporating newer observations and supporting continuous action streaming.

3.1 Rolling Denoising for Flow-Based VLAs

  • Rolling denoising updates the full H-step buffer under the latest observation, executes its refined front block, shifts remaining blocks forward, and appends fresh noise.Staggered block flow times keep the horizon in a shared action context, and steady-state emission begins after K−1 warm-up iterations.Source: §3.1 · pp. 3–4
  • Each executed block is refined across successive observations, while shared long-horizon context helps preserve mode consistency and smooth transitions across block boundaries.The method contrasts this progressive refinement with RTC-style inpainting, which stitches a newly generated suffix to a committed prefix but does not prevent a suffix from selecting a different action mode.Source: §3.1 · p. 4

3.2 Dual Decoupling for High-Throughput Deployment

  • Dual Decoupling separates sensing and VLM prefix computation from DiT denoising and action execution to recover input and output throughput while retaining rolling-buffer refinement.It trades a small amount of conditioning staleness for throughput; the selected-observation interval is set by block size S.Source: §3.2 · pp. 4–5

3.3 Staircase Training for Rolling Denoising

  • Staircase training supervises deterministic flow times tied to rolling-buffer positions, aligning training with inference and covering all denoising stages required by the scheduler.The paper reports faster convergence and lower open-loop prediction error for this structured multi-timestep supervision.Source: §3.3 · p. 5

4 Experiments

Across simulation, real-world manipulation, and specialized dynamic-control tasks, the experiments compare REACT with chunked and asynchronous baselines on success, smoothness, and reactivity. They also evaluate the training recipe’s optimization efficiency.

4.1 Experimental Setup

  • The evaluation covers RoboTwin 2.0 and six real-world platform-task pairs, with success and mean jerk as primary metrics.Policies use π0.5 variants and REACT with block size S=10 and K=5; real-world tasks include shared bowl-stacking and cap-unscrewing tasks plus platform-specific hanging tasks.Source: §4.1 · pp. 5–6, Figure 2 · p. 6

4.2 General Task Results

  • Naive shorter-horizon replanning improves observation refresh but reduces success and increases jerk; failures include repeated retries and OOD trapping.H50E10 and H10E10 also weaken long-horizon coherence, while H50E50 is less responsive to dynamic scene changes.Source: §4.2 · pp. 6–7
  • REACT w/o DD reaches 49.71% Clean and 20.86% Randomized simulation success, plus 65.7% real-world success, the highest among synchronous baselines.Its 1064.7 simulation jerk is below H50E10’s 1463.7 and H10E10’s 1468.8; it refines future blocks across observations in a full H=50 buffer.Source: §4.2 · p. 7
  • Full REACT raises Clean/Randomized simulation success from 33.29%/10.57% to 44.86%/19.86% and real-world success from 17.7% to 64.8% versus Async+RTC.It also reduces jerk from 2022.2 to 1219.0 in simulation, 284.1 to 90.1 on ARX X5, and 31.6 to 28.6 on Franka.Source: Table 1 · p. 6, §4.2 · p. 7, Table 2 · p. 7

4.3 Specialized Task Results

  • Full REACT achieves 63% Pour Rice success and 734 ms Reaction Game latency, versus 1505 ms for H50E50 and 747 ms for Async+RTC.The tasks test responses to continuously changing physical states and sudden visual events, respectively.Source: §4.3 · p. 8, Figure 4 · p. 8, Table 3 · p. 8

4.4 Training Efficiency Analysis

  • REACT reaches 52.7% average success after 3k training steps versus π0.5’s 25.0%, and matches its 30k-step performance by 15k steps.Across Click Bell, Move Playing Card Away, and Press Stapler, REACT also lowers training loss faster.Source: Figure 5 · p. 8, §4.4 · p. 9

5 Conclusion

REACT uses a persistent rolling action buffer to refine future blocks under successive observations while retaining long-horizon context. With Dual Decoupling and matched staircase training, it improves the success–smoothness–reactivity trade-off in simulation and real-world tasks.

5 Conclusion

  • REACT progressively refines future action blocks under successive observations in a persistent buffer, preserving temporal coherence across long-horizon action chunks.The design shifts away from choosing between stale long chunks and unstable frequent replanning.Source: §5 · p. 9
  • Together with Dual Decoupling and matched staircase training, REACT improves the success–smoothness–reactivity trade-off on RoboTwin 2.0 and real-world bimanual manipulation and dynamic-control tasks.Source: §5 · p. 9
  • The results suggest evaluating real-time VLA control by how effectively fresh observations enter coherent executable actions, not by inference latency alone.Source: §5 · p. 9

6 Limitations

The evaluation is limited to the flow-based π0.5 backbone, and each policy is trained for one fixed rolling schedule. The paper also reports no independent policy seeds, so statistical comparisons reflect rollout sampling noise rather than training-seed robustness.

6 Limitations

  • REACT is validated only on the flow-based π0.5 backbone, leaving adaptation to other VLA models untested.The authors note that the interface assumptions are not tied to model scale or a specific backbone, but cross-model compatibility remains unvalidated.Source: §6 · p. 9
  • Each policy is trained for one fixed rolling schedule (K, S), and the paper neither evaluates nor claims zero-shot switching between schedules.The same (K, S) = (5, 10) is used in all main experiments without per-task tuning.Source: §6 · p. 9
  • Because independent policy seeds were not trained, statistical comparisons reflect evaluation-rollout sampling noise rather than training-seed robustness.Source: §6 · p. 9

A Failure Case Analysis

The analysis attributes baseline failures to a coherence–freshness conflict: long open-loop chunks become stale, while short or asynchronous execution can weaken action consistency. REACT combines fresh observations with a persistent long-horizon buffer, and dual decoupling raises throughput without the same loss of success and smoothness.

A Failure Case Analysis

  • Across baselines, long open-loop execution becomes stale, shortened horizons weaken action-distribution support, and asynchronous execution introduces trajectory-stitching artifacts.The analysis relates these recurring mechanisms to paired success, latency, and jerk measurements.Source: A Failure Case Analysis · p. 14
  • H50E50 preserves action-mode consistency and achieves the lowest average simulation jerk (658.9), but its long open-loop execution becomes stale when scene state changes within the chunk.In Pour Rice it often continues pouring after the rice level changes, yielding 3% success; in Reaction Game latency is 1505 ms, 800 ms slower than the human teleoperation average.Source: A Failure Case Analysis · p. 14, Figure 6 · p. 15
  • H50E10 improves observation freshness but short executed prefixes can fail to complete contact-rich state changes, prompting repeated approach or grasp attempts and fragmented motion.It falls from 37.43%/11.43% to 33.71%/8.71% on RoboTwin Clean/Randomized settings and from 58.7% to 35.9% on real robots; Bottle Cap Unscrewing success is 0% on both ARX X5 and Franka.Source: A Failure Case Analysis · pp. 14–15, Figure 8 · p. 16, Figure 7 · p. 16
  • H10E10 removes the training–execution horizon mismatch but loses long-horizon context, with success falling to 10.86%/2.43% on RoboTwin Clean/Randomized settings and 21.0% on real robots.On Franka it reaches 0% on all three general tasks.Source: A Failure Case Analysis · p. 15
  • Async+RTC raises throughput by hiding inference time, but stitched chunks can produce abrupt corrections and high jerk, alongside only 17.7% average real-world general-task success.Average simulation jerk is 2022.2 versus 1219.0 for full REACT; ARX jerk is 284.1 versus 90.1.Source: A Failure Case Analysis · pp. 15, 17
  • 65.7% vs. 35.9%, 21.0%, and 17.7% real-world success for REACT w/o DD versus H50E10, H10E10, and Async+RTC; full REACT has lower jerk than Async+RTC in simulation and on ARX X5.Simulation jerk is 1219.0 vs. 2022.2, and ARX jerk is 90.1 vs. 284.1.Source: A Failure Case Analysis · p. 17

B Throughput Comparison

  • Throughput analysis defines total inference time as Tinfer = TVLM + N · TDiT and selected-observation rate as fcam/S for REACT variants with block size S = H/K.H is the predicted action horizon, He the executed horizon, and N the number of denoising steps.Source: B Throughput Comparison · p. 17

B.1 Qualitative Analysis

The section contrasts sequential, asynchronous, and accelerated inference with REACT’s decoupled pipelines, which separate sensing, VLM encoding, denoising, and execution to improve observation and action throughput.

B.2 Summary of Throughput Formulas

  • Synchronous inference completes VLM encoding and N denoising steps before executing H_e actions, with VLM encoding typically dominating the cycle.The throughput formulas are summarized in Table 4, with inference time expressed as T_infer = T_VLM + N · T_DiT.Source: §B.1 · pp. 17–18, §B.2 · p. 19, Table 4 · p. 20

B.1 Qualitative Analysis

  • Asynchronous inference overlaps next-chunk inference with execution but still processes only one observation per cycle; acceleration helps synchronous throughput but adds no asynchronous throughput when inference is already hidden.REACT without Dual Decoupling instead uses fresh observations at each denoising step and emits S=H/K actions, with encoding, denoising, and execution remaining sequential.Source: §B.1 · p. 18
  • Dual Decoupling runs sensing, VLM encoding, DiT denoising, and action execution as separate streams, using fresh completed embeddings and the rolling action buffer.The scheduler selects observations every S camera periods and commits the next S executable actions.Source: §B.1 · p. 18, Figure 9 · p. 19
  • When the worker pool and DiT sustain selected observations, REACT streams executable actions at the camera cadence; at 30 Hz and S=5, this is 6 Hz observation processing and 30 Hz action output.The selected-observation rate is below the approximately 27 Hz single-worker reference throughput, and the DiT ceiling is about 333 Hz at 3 ms per update.Source: §B.1 · p. 19

S TVLM + TDiT + S/fctrl

The section analyzes throughput trade-offs, documents the hardware and model configuration, and describes REACT’s rolling inference and training alignment.

S T_VLM + T_DiT + S/f_ctrl

  • A 2× inference speedup raises synchronous H_e=50 input throughput only from 0.58 Hz to 0.59 Hz, while shortening execution to H_e=5 raises Φ_in to 4.28 Hz but lowers Φ_out to 21.4 Hz.For H_e=10, asynchronous execution raises throughput from 2.50 Hz/25.0 Hz to 3.00 Hz/30.0 Hz, but future chunks can be conditioned on observations mismatched to execution state.Source: S TVLM + TDiT + S/fctrl · pp. 20–21

C.1 Hardware Platforms

  • The real-robot evaluation uses Acone/ARX X5 and dual-arm Franka Research 3 platforms, with three 30 Hz RGB cameras and 30 Hz robot control on each described setup.The X5 has two 6-DoF arms with 1-DoF grippers; the Franka setup uses two 7-DoF arms and Cartesian impedance control.Source: §C.1 · p. 21

C.2 Model Architecture

  • All variants use π0.5 with a PaliGemma-3B vision-language model and 300M-parameter action expert, while REACT uses one Euler denoising step per update on the H-step buffer.The action representation is 14-dimensional joint space on X5 and 20-dimensional end-effector space on Franka; the H=50 horizon is partitioned into K blocks of size S.Source: §C.2 · p. 22

C.3 Training Configuration

  • Training hyperparameters for the simulation and real-robot policies are summarized in Table 6.The supplied passage identifies the table's scope but does not give individual hyperparameter values.Source: §C.3 · p. 22, Table 6 · p. 22

C.4 Inference Configuration

  • At episode reset, REACT initializes the buffer from the robot reset state with blockwise staircase noise, then performs K−1 warm-up updates before emitting evaluated actions.After warm-up, each control iteration emits one S-action block using the latest available observation; full REACT's asynchronous VLM workers publish embeddings to a latest-ready cache.Source: §C.4 · pp. 22–23

C.5 Train-Inference Consistency

  • Staircase training aligns each block's position and flow time with inference: blocks advance one Euler step, shift toward execution, and enter the front block at the clean endpoint.A block enters at the tail as pure noise, receives observation-conditioned updates across K stages, and exits at the front; training supervises the intermediate stages.Source: §C.5 · pp. 23–24

C.6 Benchmarks and Evaluation Protocol

The evaluation protocol covers seven RoboTwin simulation tasks and real-world tasks on two bimanual platforms, with success, trajectory smoothness, and specialized-task measures.

C.6 Benchmarks and Evaluation Protocol

  • Seven representative RoboTwin 2.0 tasks are evaluated in Clean and Randomized conditions using 50 demonstrations per task and 100 rollouts per task-condition pair.The selected tasks cover tool use, precision contact, grasping, card manipulation, and articulated-object manipulation.Source: §C.6 · p. 24
  • Real-world evaluation uses six platform-task pairs across ARX X5 and Franka R3, with 30 trials per task under shared randomized object placements.The tasks span size-based reasoning, bimanual coordination, and deformable or articulated-object manipulation.Source: §C.6 · p. 24

C.7 Evaluation Metrics

  • Success rate is the percentage of completed trials, while trajectory smoothness is mean jerk norm and is averaged only over successful episodes.Lower jerk indicates smoother motion; specialized evaluations report Pour Rice success and Reaction Game stimulus-to-click latency, with a human teleoperation baseline using the same timing path.Source: §C.7 · p. 24

D Full Trajectory Smoothness Results

The section reports demonstration jerk references and compares the full-execution-horizon policy and full REACT with asynchronous execution plus RTC.

D Full Trajectory Smoothness Results

  • Demonstration trajectories average jerk values of 570.1 in simulation, 41.9 on ARX X5, and 15.3 on Franka.The full-execution-horizon H50E50 policy is closest to these reference values.Source: D Full Trajectory Smoothness Results · p. 25
  • Full REACT reduces jerk relative to Async+RTC on the real-robot platforms and in simulation.Table 8 reports mean jerk norm by task and platform, with lower values indicating smoother motion.Source: D Full Trajectory Smoothness Results · p. 25, Table 8 · p. 25
  • Trajectory smoothness results are reported as mean jerk norm of executed joint trajectories, with best and second-best values marked in the table.The supplied text does not provide the per-task cell values.Source: Table 8 · p. 25

E Ablation Studies

The ablation varies block length S and rolling-block count K, which govern observation-update frequency and multi-observation refinement. Results favor (S, K) = (10, 5) within the admissible pretrained-horizon grid, though the best granularity varies by task.

Ablation parameters

  • Block length S sets how often the policy receives a new observation and emits an action block, while K determines how many rolling refinements and recent observations shape each executed block.Each block appended as tail noise is refined once per rolling update before execution after K staircase stages.Source: E Ablation Studies · p. 25

Scope of the evaluated S×K grid

  • The evaluated grid is restricted to K·S ≤ 50 because the pretrained π0.5 checkpoint has a fixed action horizon H=50; larger horizons require full VLA re-pretraining.S=10, K=5 saturates that horizon and permits the largest multi-observation convergence budget it allows.Source: E Ablation Studies · pp. 25–26

Aggregate trend

  • At S=10, mean RoboTwin clean success rate is 49.7% for K=5, versus 38.0% for K=4, 34.9% for K=3, and 19.9% for K=2.Short-block configurations do not surpass S=10, K=5; the paper selects it for the final configuration.Source: Figure 11 · p. 26, E Ablation Studies · p. 26

Schedule sensitivity and practical selection

  • At fixed S=10, mean success rises from 13.6% to 49.7% as K grows from 1 to 5; at fixed K=5, it rises from 6.3% to 49.7% as S grows from 1 to 10.The same (K, S) = (5, 10) setting is used in all main experiments and is optimal or tied for optimal on 6 of 7 RoboTwin tasks, except turn_switch.Source: E Ablation Studies · p. 26
  • The proposed selection recipe is to choose S for the required reaction timescale and compute budget, then use the largest K satisfying K·S ≤ 50.Smaller S requires more frequent VLM and DiT updates; the single denoising update per control cycle is not independently tuned.Source: E Ablation Studies · p. 26

Per-task trend

  • The best block granularity depends on manipulation structure, with Figure 12 showing per-task success across S and K.The heatmaps use a shared 0–100% color scale, and hatched cells mark configurations not evaluated.Source: E Ablation Studies · p. 26, Figure 12 · p. 27

F OOD Dataset Action Prediction Open-Loop MSE

On three demo_randomized RoboTwin tasks, REACT’s final training checkpoints have lower average open-loop action-prediction MSE than the π0.5 baseline. The appendix also presents a qualitative prediction comparison for one Move Playing Card Away example.

Evaluation setup

Aggregate OOD prediction

Qualitative prediction

  • Figure 14 presents two visualizations of one demo_randomized Move Playing Card Away example, comparing predicted action sequences from the baseline and REACT with the demonstration trajectory.The passage identifies the comparison but does not state which prediction is closer.Source: F OOD Dataset Action Prediction… · p. 27, Figure 14 · p. 28

Related papers

Loading 2610.12007v1…