PolicyTrim cuts VLA robot task time by making action chunks longer and trajectories shorter

新智元 · wechat · 2026-07-21

A new paper proposes **PolicyTrim**, a post-training method for VLA robots that improves real-world execution efficiency without changing the architecture or retraining from scratch. - The core problem is not just inference latency, but **policy efficiency**: how many reliable action chunks and physical steps are needed to finish a task. - PolicyTrim uses a two-stage RL-style post-training pipeline: - **Reliable Action Chunk Extension**: expands the part of an action chunk that can be safely executed. - **Redundancy-Aware Step Reduction**: rewards shorter successful trajectories and suppresses wasted corrective motions. - Reported gains include **3× action-chunk utilization**, up to **51.4% fewer physical steps**, and up to **5.83× end-to-end speedup** on LIBERO with π0.5, while keeping success rate above **98%**. - The method was also tested on **ManiSkill**, **Meta-World**, and real robot tasks on an **Agilex Piper** arm, where it delivered about **1.86×** speedup in the standard setup. - The paper’s key claim is that making robots “take fewer wrong steps” is as important as making each inference faster, and the two approaches can stack.

Original post →

More from Embodied

Embodied channel →