PiHand: latent hand-object interaction priors cut dexterous-hand RL training data 10x
KyleMorgenstein · x · 2026-11-14
Researchers from University of Tübingen, MPI Informatics, and UC Berkeley present PiHand, a physics-grounded framework tackling the huge sample requirements of RL for dexterous hands.
The method works in three stages:
- Train expert policies in simulation that track mocap demonstrations
- Distill them into a unified hand-object interaction prior via a goal-conditioned VAE
- Run RL with bounded residual exploration inside the latent space, using a single object-tracking reward and no curriculum
Results:
- Matches the best baseline with 10x less training
- Zero-shot generalization to unseen objects and trajectories (validated on GRAB and DexYCB)
- Achieves catching and bimanual handovers that barely appear in training data
Target applications span VR/AR, gaming, and robotic manipulation.
More from Embodied
- The $50 Bee wearable is crushing $700 AI devices, says new analysis — petrusenko_max · 2026-11-14
- LabMate robot demo: Gemma 4 on a reBot arm fetches missing lab tools — agihouse_org · 2026-11-14
- Gemma hackathon winner: robot brain failover to another machine in 625ms, no cloud — agihouse_org · 2026-11-14
- Cybertruck police car spotted at Starbase with Starship in the background — shivon · 2026-11-14
- OpenAI insider: robotics is "alive and well", with 15 open roles hiring — ashwinl · 2026-11-14
- Teleoperating Mundane's robot via force-feedback exoskeleton: can feel empty vs full bottle — chris_j_paxton · 2026-11-14