Demystifying Action Space Design for Robotic Manipulation Policies

A research team from unnamed institutions, led by Yuchun Feng, has published a significant paper on arXiv titled 'Demystifying Action Space Design for Robotic Manipulation Policies.' The work addresses a critical but often overlooked component in robotic imitation learning: the fundamental design of the action space itself. While recent progress has focused on scaling data and model capacity, the choice of how a policy outputs actions—its action space—has remained guided by legacy heuristics. This study aims to bring scientific rigor to this foundational design decision, which shapes the entire optimization landscape for policy learning.

The researchers conducted a large-scale empirical study involving over 13,000 real-world rollouts on a bimanual robot and the evaluation of more than 500 trained models across four manipulation scenarios. They dissected the action design space along temporal (absolute vs. delta actions) and spatial (joint-space vs. task-space) axes. The key results provide clear, data-backed guidance: policies designed to predict delta actions (changes from the previous state) consistently outperform those predicting absolute positions. Furthermore, the choice between joint-space (controlling robot joints directly) and task-space (controlling end-effector pose in the world) involves a trade-off. Joint-space representations favor control stability, while task-space representations offer better generalization to new situations. This work provides a structured framework for engineers to make informed design choices, potentially accelerating the development of more reliable and capable robotic manipulation policies.