Contact-rich manipulation requires a robot to coordinate not only where to move, but also how contact and force should evolve during an interaction. The authors of a new arXiv paper observe that most robot policies represent actions primarily as motion trajectories, which may leave this interaction dimension under-specified.
To address this, they propose representing actions as 'interaction latents'—a latent action space capturing the evolving interaction between robot and environment, rather than just the kinematic path. This shifts the policy's output from motion trajectories to the desired contact and force progression.
The abstract does not describe experimental details, so this digest cannot assess empirical results. The conceptual shift is noteworthy nonetheless: if interaction itself is the true object of control in contact-rich manipulation, then the action representation should reflect that.