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

arXiv:2610.12007 · cs.RO, cs.AI · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa, and with me are Dev and Taro, guest researcher.

Dev: Today's paper: "REACT: Rolling Denoising and Dual Decoupling for Reactive Robot Control with VLA Models".

Rosa: The gist REACT reformulates flow-based action generation into a receding-horizon rolling denoising process to make flow-based VLAs more reactive while preserving long-horizon context How it works REACT introduces a rolling-denoising…

Dev: First, who's behind it and why it matters.

Title and authors: Rosa: So we're looking at this paper called "REACT: Rolling Denoising and Dual Decoupling for Reactive Robot Control with VLA Models" by Houlong Xiong, Zhenqi Qiu, Zechen Wang, Suohang Zhang, Yiyu Ren, Wanting Xu, Hongfei Niu, Chengyang He and Ge Sun.

Dev: Yeah the title itself tells you a lot about what they're trying to do. It's about rolling denoising and dual decoupling for reactive robot control with VLA models.

Rosa: Basically they're taking these flow-based action generation methods, which are good for making smooth actions over time, and turning them into something more reactive by using a rolling denoising process instead of just generating chunks from scratch every single time.

Taro: So the core idea is to keep a persistent action buffer with staggered flow times instead of regenerating each chunk from scratch at every step. That sounds like it tries to solve that problem where new observations don't immediately influence the future actions inside a chunk, which is a big issue for robots in dynamic places.

Dev: Exactly. They suggest that at each control step, they denoise the full horizon using the latest observation to get the cleanest action block, execute that part, shift the remaining future blocks forward, and then add fresh noise to the tail.

Rosa: It’s trying to make sure each executed action block is refined across multiple recent observations before it goes out into the world. That should help with consistency while keeping things up-to-date.

The paper's summary: Rosa: Now let's talk about what they actually found in the REACT paper. They introduced this rolling denoising framework that maintains a persistent action buffer with staggered flow times instead of regenerating each chunk from scratch.

Dev: The key mechanism is that at every control step, they take the latest observation and use it to denoise the full horizon, then execute the cleanest action block while shifting forward the rest of the buffer. This keeps things moving smoothly.

Taro: They use a per-position staircase time schedule where each flow time τj is defined as floor j over S plus one, which gives them a specific way of handling those action chunks across different positions. That seems like a structured way to manage the temporal aspect of the problem.

Rosa: The training part is also important because they use matched staircase training where they train the action expert on those exact same per-position flow times used during their rolling inference. This means the loss supervises those K deterministic denoising stages in one example, which aligns it perfectly with how they deploy it.

Dev: That alignment is what allows the model to learn this deployed action-flow schedule without needing a ton of extra optimization steps for every little change. It makes learning the schedule much more efficient.

Taro: So what this means for us is that we get a system that learns to follow a specific, structured way of refining its actions in real-time based on what it sees right now.

The paper's improvements: Rosa: The authors highlight several specific improvements they made to the original flow-based VLA approaches. They focus on fixing that tension between coherence and reactivity.

Dev: They introduce dual decoupling which separates sensing, VLM encoding, DiT denoising, and action execution without changing any of the model weights. This is crucial because it allows them to handle high-frequency updates while keeping things decoupled in terms of processing stages.

Taro: The dual decoupling lets them select one observation every S camera periods to get a selected-observation rate of fcam over S, and they have a pool of M VLM workers encoding frames asynchronously into a cache. This is designed to recover the input and output throughput that other methods lose when you try to serialize everything.

Rosa: The training alignment is another major point. They use matched staircase training where they supervise the K deterministic denoising stages in one example with this loss: LREACT equals one over K times the sum from zero to S of X minus one k equals zero one over S times the sum from j in B k of v(j) theta minus u(j) τj squared.

Dev: That loss is specifically designed to cover the time-position pairing used by rolling inference, because block k is always queried at flow time (k plus one) over K before shifting one position toward execution, so the training loss supervises the same blockwise noise levels used in deployment.

Taro: This means they’re not just learning a general policy; they’re learning the specific action-flow schedule that their scheduler will use when it's actually running. It tightens the loop between training and inference significantly.

Conclusion: Rosa: So to wrap up, REACT is essentially a rolling-denoising framework that makes flow-based VLA policies more reactive while keeping the long-horizon context consistent. They use dual decoupling and matched staircase training to get there.

Dev: The main implication for us is that they've managed to improve the success and smoothness trade-off on RoboTwin two point zero, showing substantial improvements over baselines like Async+RTC, getting real-world success up from seventeen point seven percent to sixty-four point eight percent.

Taro: For reaction games specifically, they achieved a mean latency of seven hundred thirty-four milliseconds on Franka R3, which is ahead of Async+RTC at seven hundred forty-seven milliseconds and only about twenty-nine milliseconds above human teleoperation at seven hundred five milliseconds.

Rosa: That shows that by coupling fast closed-loop correction with smoother action evolution, you can react quickly while still producing the coherent actions needed for physical execution, especially in tricky tasks like Pour Rice.

Dev: The throughput analysis also says that for full REACT, the output throughput reaches the camera cadence whenever the VLM pool and DiT keep up. The system processes observations at fcam over S and emits S to the power of four outputs per update.

Taro: Their limitation is that they only validated REACT on this specific flow-based VLA backbone π0 point 5, meaning we haven't seen how it adapts to other types of VLA models yet. That’s something they need to work on next.

Rosa: Well, REACT is a rolling-denoising framework for making flow-based VLA policies more reactive while preserving the temporal coherence of long-horizon action chunks. We've covered the basics of how it works and what the results are.

Dev: The dual decoupling and matched staircase training are key to getting that real-time deployment working efficiently under compute constraints.

Taro: It really shows how structuring the training around the inference schedule can lead to much faster learning for these kinds of policies.

Shanghai Jiao Tong University 2 PrimeBot 3 ShanghaiTech University 4 Zhejiang University 5 National University of Singapore

cs.RO, cs.AI

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: Accepted to CoRL 2026 as Spotlight. Project page: https://react-vla.github.io

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 85/100

The gist: The gist REACT reformulates flow-based action generation into a receding-horizon rolling denoising process to make flow-based VLAs more reactive while preserving long-horizon context How it works

Key concepts

Flow-based VLA Models
These models generate robot actions by creating continuous motion chunks over time. The challenge is balancing long, smooth actions with the need for quick adjustments when things change in the environment.
Rolling Denoising Framework
Instead of starting a new action plan from scratch every time, this method continuously refines the current action block using the newest sensor data. It maintains a persistent buffer and updates it incrementally, ensuring that each executed part of the motion is refined across several recent observations.
Dual Decoupling
This technique separates different parts of the robot control system—sensing, language modeling (VLM), denoising (DiT), and action execution. This separation allows the system to update observations at a high frequency while managing computational resources efficiently for real-time performance.

Terminology

Summary

The gist REACT reformulates flow-based action generation into a receding-horizon rolling denoising process to make flow-based VLAs more reactive while preserving long-horizon context

How it works

REACT introduces a rolling-denoising framework that maintains a persistent action buffer with staggered flow times instead of regenerating each chunk from scratch 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

The framework utilizes a per-position staircase time schedule where the flow time for block j is defined as τj = ⌊j/S⌋ + 1, which yields τ = [1/K,..., 1/K S, 2/K,..., 2/K S,..., 1,..., S]. Given the latest observation oi, one DiT pass predicts v(i) = vθ(x(i), τ, oi), and one Euler update advances every block by one denoising stage: xe(i) = x(i) − 1/K v(i) After the Euler step, REACT executes the updated front block, shifts the remaining blocks left, and appends fresh Gaussian noise to the tail: x(i+1) = h Be(i) 1, Be(i) 2,..., Be(i) K−1, ϵnew i This rolling denoising structure improves observation freshness while preserving the long-horizon mode consistency of action chunking

Dual Decoupling for High-Throughput Deployment

To support real-time control, REACT introduces dual decoupling, which separates sensing, VLM encoding, DiT denoising, and action execution without changing model weights This inference-time scheduler selects one observation every S camera periods to yield a selected-observation rate of fcam/S A pool of M VLM workers encodes selected frames asynchronously into a latest-ready cache, and each DiT update consumes the freshest completed embedding together with the rolling buffer The purpose of this decoupling is to recover the input and output throughput that REACT would otherwise lose to serialized synchronization

Training Alignment

REACT employs matched staircase training, where it trains the action expert on the same per-position flow times used by rolling inference The loss supervises the K deterministic denoising stages in one example: LREACT = 1/K Σ X−1 k=0 1/S Σ X j∈Bk v(j) θ − u(j) τj 2 This objective covers the time-position pairing used by rolling inference: block k is always queried at flow time (k+1)/K before shifting one position toward execution, so the training loss supervises the same blockwise noise levels used by the deployed scheduler

Experimental Findings

Across RoboTwin 2.0 simulation, real-world bimanual manipulation on ARX X5 and Franka R3, and targeted Pour Rice and Reaction Game evaluations, REACT improves task success and reduces reaction latency while producing smoother trajectories than frequent-replanning and asynchronous baselines Full REACT achieves the highest Pour Rice success rate (63%) by combining fast closed-loop correction with smoother action evolution, avoiding both stale overpouring and short-horizon mode oscillations Full REACT substantially outperforms Async+RTC: it improves Clean/Randomized simulation success from 33.29%/10.57% to 44.86%/19.86%, and real-world success from 17.7% to 64.8%

Throughput Analysis

The throughput formulas show that for full REACT, the output throughput reaches the camera cadence whenever the VLM pool and DiT keep up The system therefore processes observations at fcam/S and emits S 4 outputs per update

Limitations

A primary limitation is that we validate REACT only on the flow-based VLA backbone π0.5 [7], leaving adaptation to other VLA models untested

Conclusion

This paper presented REACT, a rolling-denoising framework for making flow-based VLA policies more reactive while preserving the temporal coherence of long-horizon action chunks Together with Dual Decoupling and matched staircase training, this design improves the success–smoothness–reactivity trade-off on RoboTwin 2.

Improvements for AI systems

  1. Bold header: Rolling Denoising for Flow-Based VLAs

The improved system maintains a persistent action buffer with staggered flow times instead of regenerating each chunk from scratch, allowing each executed action block is refined across multiple recent observations before deployment. This ensures that adjacent blocks are refined under a shared long-horizon action context, keeping them coupled in the same action mode and smoothing transitions across block boundaries.

  1. Bold header: Dual Decoupling for High-Throughput Deployment

The system implements dual decoupling, which separates sensing, VLM encoding, DiT denoising, and action execution, enabling high-frequency observation updates and action streaming under practical compute constraints. This allows the system to achieve throughput bounded by Φin = min fcam/S, 1/TDiT while preserving the quality benefits of rolling denoising.

  1. Bold header: Staircase Training for Rolling Denoising

The training process is modified to supervise K deterministic denoising stages in one example, which align[s] the flow supervision with rolling inference. This results in a model that learns the deployed action-flow schedule with fewer optimization steps, leading to faster convergence and lower open-loop prediction error.

  1. Bold header: Improved Dynamic Control via Coherence Preservation

The system addresses the coherence–freshness trade-off by using reactive refinement while preserving context, as evidenced by its ability to improve task success rates over baselines like H50E10 and Async+RTC. This allows the robot to react quickly while preserving the smooth, temporally coherent actions needed for physical execution, particularly in dynamic tasks like Pour Rice.

  1. Bold header: Enhanced Reaction Latency Reduction

For reaction games, the system significantly reduces latency by achieving a mean latency of 734 ms on Franka R3, ahead of Async+RTC (747 ms) and only 29 ms above human teleoperation (705 ms), by coupling fast closed-loop correction with smoother action evolution.

  1. Bold header: Optimized Training Efficiency

The staircase objective leads to faster learning, as REACT reaches 52.7% average success after only 3k steps versus 25.0% for the π0.5 baseline and matches the baseline’s 30k-step performance using only 15k steps. This reduces the computational cost required to train a policy capable of real-time deployment.

Abstract

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.

Sources

Related papers