SPACE framework lets robots learn from each other across different bodies

A team from KAIST (Haeone Lee, Byeongguk Jeon, Suchae Jeong, Jian Kim, Kimin Lee) has published SPACE, a framework designed to enable robot learning from cross-robot data toward generalist policies. The core insight is that robot actions are tied to the dynamics of the collecting robot, making it hard to train a single policy across different hardware. SPACE solves this by using Cartesian state delta – the geometric displacement of the end-effector – as a universal action representation. This decouples the desired motion from the robot's specific dynamics.

The framework has two components: (1) a Cartesian state delta policy that predicts the desired end-effector displacement, and (2) an Action Adapter that converts that displacement into robot-specific control commands. SPACE addresses variation at three levels: across different robot embodiments, across hardware units of the same embodiment, and within a single robot during operation. Experiments show SPACE substantially outperforms policies that directly predict control commands when learning from data across different robots. It also remains robust under runtime dynamics shifts like changes in control frequency, object weight, or controller gains. Project page and code are available.