cross-embodiment-learning

The paper introduces a method for extracting topological necessities as mechanism-invariant subgoals from offline trajectories, enabling cross-embodiment transfer in goal-conditioned reinforcement learning, with empirical improvements on benchmark tasks.