X-Reset: Scaling Object-Centric Reinforcement Learning via Cross-Embodiment Resets
X-Reset uses retargeted human hand-object states as RL resets to train cross-embodiment dexterous manipulation policies.
X-Reset addresses exploration in object-centric reinforcement learning by kinematically retargeting human hand-object demonstrations into noisy robot states, filtering unstable states, and sampling the rest as training resets. Policies use general-purpose object-centric rewards and depend only on object state and goal, not imitation of human motion. The method trains generalist policies on 20 objects across a 22-DoF hand on two arms and a parallel-jaw gripper, generalizes to unseen objects, tolerates imperfect hand-pose estimates, and transfers zero-shot from simulation to real robots.
- Human demonstrations enter training only through the reset distribution.
- Policies trained on 20 objects across three robot embodiments.
- Scales with object count and generalizes to unseen objects.
- Behaviors transfer zero-shot from simulation to real robots.
From the original article
1 min read at arxiv.orgReinforcement learning (RL) in simulation can train dexterous manipulation policies without robot demonstrations, but training a single generalist policy with task-agnostic rewards faces a severe exploration problem: approaching, grasping, and reorienting diverse objects with many degrees of freedom is difficult to discover from scratch.
By Prithwish Dan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.