X-Reset: Scaling Object-Centric Reinforcement Learning via Cross-Embodiment Resets
The paper uses human hand-object states as reset points so robot RL can skip the hardest cold-start exploration.
X-Reset retargets those hand-object states into noisy robot states, filters out unstable ones in simulation, and samples the rest during training. The policy is trained with general-purpose object-centric rewards and depends only on object state and goal, not on tracking the demonstrations. The authors report generalist policies across 20 objects and three robot embodiments, including a 22-DoF hand on two arms and a parallel-jaw gripper. They also claim scaling with more objects, generalization to unseen objects, tolerance of imperfect hand-pose estimates, and zero-shot sim-to-real transfer. ArXiv · AI/CL/LG's note
X-Reset retargets those hand-object states into noisy robot states, filters out unstable ones in simulation, and samples the rest during training. The policy is trained with general-purpose object-centric rewards and depends only on object state and goal, not on tracking the demonstrations. The authors report generalist policies across 20 objects and three robot embodiments, including a 22-DoF hand on two arms and a parallel-jaw gripper. They also claim scaling with more objects, generalization to unseen objects, tolerance of imperfect hand-pose estimates, and zero-shot sim-to-real transfer. ArXiv · AI/CL/LG's note
score 5