Latent RL Prior for dexterous manipulation

A two-stage framework reuses control knowledge to learn visual policies from limited demonstrations.

Latent RL Prior is a two-stage framework. It reuses control knowledge for data-efficient visual policy learning.

During privileged pretraining, the system distills large-scale reinforcement learning (RL) rollouts. It uses a flow-matching action decoder conditioned on proprioception and latent task context.

Proprioception is used for conditioning the decoder. A latent task context is also used for conditioning.

During visual adaptation, the decoder is frozen. A visual interface maps images to the latent context using action supervision alone.

This method outperforms full-parameter fine-tuning of π0.5\pi_{0.5} and GR00T N1.7 in the low-data regime. In door lifting, it uses 10×10\times fewer demonstrations. In bimanual assembly, it uses 5×5\times fewer demonstrations.

The decoder can be shared across arms. It learns coordinated visual contexts for bimanual manipulation.

Key facts

  • Latent RL Prior outperforms π0.5\pi_{0.5} and GR00T N1.7 in low-data regimes.
  • The method uses 10×10\times fewer demonstrations for door lifting.
  • The method uses 5×5\times fewer demonstrations for bimanual assembly.

All sections

Edition