Demystifying Action Space Design for Robotic Manipulation Policies

The specification of the action space plays a pivotal role in imitation-based robotic manipulation policy learning, fundamentally shaping the optimization landscape of policy learning. While recent advances have focused heavily on scaling training data and model capacity, the choice of action space remains guided by ad-hoc heuristics or legacy designs, leading to an ambiguous understanding of robotic policy design philosophies. To address this ambiguity, we conducted a large-scale and systematic empirical study, confirming that the action space does have significant and complex impacts on robotic policy learning. We dissect the action design space along temporal and spatial axes, facilitating a structured analysis of how these choices govern both policy learnability and control stability. Based on 13,000+ real-world rollouts on a bimanual robot and evaluation on 500+ trained models over four scenarios, we examine the trade-offs between absolute vs. delta representations, and joint-space vs. task-space parameterizations. Our large-scale results suggest that properly designing the policy to predict delta actions consistently improves performance, while joint-space and task-space representations offer complementary strengths, favoring control stability and generalization, respectively.

Robots learn many manipulation skills by imitating demonstrations, such as picking up objects or moving items around. While recent robot learning research has focused on using larger models and more training data, an important design choice has received much less attention: how robot actions should be represented. There are many ways to describe a robot’s actions. For example, a robot can predict how much each motor should move, or instead predict where its hand should move in space. Actions can also describe absolute targets or small changes relative to the current position. These choices strongly affect how easily robots learn and how reliably they perform, yet they are often chosen based on convention rather than evidence. In this work, we conduct a large-scale study of robot action representations using more than 13,000 real-world robot trials and over 500 trained models. We compare different ways of representing robot actions across multiple tasks and training settings. Our results show that predicting relative changes in movement consistently improves robot performance. We also find that different action representations have different strengths: some lead to more stable control, while others generalize better to new situations. Overall, our study shows that action representation is a key factor in robot learning and should be studied more carefully when designing future robotic systems.