BackAI Era

AI Era

Embodied AI
2026-07-21 09:01:08

PolicyTrim targets VLA robot efficiency with up to 5.83x end-to-end speedup

A team led by Professor Lei Yinjie at Sichuan University has introduced PolicyTrim, a two-stage post-training framework designed to improve the deployment efficiency of Vision-Language-Action, or VLA, robots without changing model architecture or recollecting expert data. The method focuses on what the paper calls policy efficiency: how many actions in a predicted chunk can be executed reliably, and how many real-world physical steps are needed to finish a task. In the first stage, PolicyTrim expands the reliable execution horizon of action chunks through dynamic execution horizon exploration. In the second, it reduces redundant physical steps with a reward tied to shorter successful trajectories, while using group-anchored regularization to avoid brittle shortcuts. The paper reports tests across LIBERO, ManiSkill, Meta-World and real-world robot tasks, covering architectures including π0.5, OpenVLA-OFT and GR00T. Results cited in the paper include a 3x increase in action chunk utilization, a 51.4% reduction in physical steps, and a peak 5.83x end-to-end speedup for π0.5 on LIBERO while keeping success rate above 98%.

290
PolicyTrim targets VLA robot efficiency with up to 5.83x end-to-end speedup