Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

What Makes a VLA Reactive? A Study of Execution Strategies for Dynamic Manipulation

Zhuoyue Li, Zhendong Han, Pankhuri Vanjani, Wenting Li, Rudolf Lioutikov

November 2026

TL;DR A study of what makes the π0.5 VLA react to moving objects, in simulation and on three real Franka tasks. How the policy is executed matters as much as the policy itself. On a real conveyor belt, asynchronous execution lifts success from 20% to 60-75%, and the best continuity-preserving method reaches 95%.

What they do

VLAs predict chunks of future actions. When objects move, a chunk can go stale before it has finished executing. This study takes one pretrained VLA, π0.5, and asks what makes it react to motion. It runs on two dynamic simulation benchmarks (DOM, DOMINO) and three real Franka Panda tasks: picking a red cube off a conveyor belt, placing a cylinder into a cup pulled on a rope, and grasping a rolling can. Seven methods from recent papers are adapted into one π0.5 pipeline.

How it works

  • Horizons. The prediction horizon is varied during fine-tuning, and the execution horizon (actions run before replanning) at deployment. Adaptive Action Chunking (AAC), which ends a chunk early when sampled actions disagree, is also tested.
  • Scheduling and continuity. Synchronous execution, where the robot waits for each new chunk, is compared with asynchronous execution. FASTER and DynamicVLA deliver actions sooner; RTC and ABPolicy keep successive chunks continuous.
  • Policy changes. PUMA adds optical-flow history and supervision from future object features. DSRL trains a latent-noise actor-critic with RL while the VLA stays frozen.

Results

  • The best prediction horizon is about one second on both benchmarks. AAC never beats the best fixed execution horizon.
  • On the real conveyor with an execution horizon of 10 steps, synchronous π0.5 succeeds in 20% of trials and asynchronous methods in 60-75%. ABPolicy is best at 95%, and RTC matches the lower-latency FASTER at 75%.
  • On the fast belt, raising the action rate from 20 to 30 Hz lifts ABPolicy from 13% to 93% and FASTER from 0% to 33%.
  • DSRL raises the targeted case from 32% to 47% on DOM and from 0% to 60% on the hardest conveyor case. Over all nine conveyor conditions it lifts the 1k checkpoint to 64-69%, but a shorter execution horizon gets the same checkpoint to 71% with no training.

Limitations

The study covers only π0.5, adapted rather than full versions of the other methods, and a small set of tasks. Real-robot comparisons use 15-20 trials per setting. The can and cup are moved by hand, so their motion is not identical across methods.

Figures

From Figure 1: Whether frequent replanning helps a VLA in dynamic manipulation depends on the prediction and executio...
From Figure 1: Whether frequent replanning helps a VLA in dynamic manipulation depends on the prediction and execution horizons, inference scheduling, chunk continuity and policy behavior.
From Figure 5: Synchronous execution pauses while the next chunk is inferred; asynchronous execution overlaps inferen...
From Figure 5: Synchronous execution pauses while the next chunk is inferred; asynchronous execution overlaps inference with motion but can cause discontinuities between chunks.
From Figure 7: The real-robot conveyor-belt picking task.
From Figure 7: The real-robot conveyor-belt picking task.
Abstract

Vision-language-action (VLA) models predict chunks of future actions, and when objects move, those chunks go stale before they are executed. We study what makes a pretrained VLA reactive, using π0.5 on two dynamic simulation benchmarks and three real-robot tasks. We vary the prediction and execution horizons, compare synchronous and asynchronous inference, test methods that deliver actions sooner or keep successive chunks continuous, and adapt the policy with motion cues and latent-space reinforcement learning. The best prediction horizon is about one second, and the best execution horizon depends on the checkpoint. Short horizons help more with asynchronous execution, and methods that keep the trajectory continuous outperform those that only reduce latency. Adapting the policy repairs targeted failures, but on the conveyor task changing the execution horizon gained as much. Reactivity depends on the execution pipeline as much as on the policy, so both should be evaluated together.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.