A new preprint on arXiv tackles the problem of training a single robotic manipulation policy that works across different robot bodies. The authors frame the core difficulty as embodiment-dependent transition dynamics: each robot moves and responds differently, so a policy trained on one platform does not naturally transfer to another.

The proposed direction is collective learning, which aggregates data from multiple robots rather than treating each embodiment in isolation. To make this aggregation work, the paper introduces predictive dynamics adaptation, a mechanism intended to account for the differences in how each robot transitions between states.

The work focuses on state-based Cartesian manipulation, where control is expressed in Cartesian coordinates rather than joint space. The abstract argues that this setting makes the embodiment mismatch especially visible, and that predictive dynamics adaptation is a promising route toward a unified manipulation policy.