Latent RL Priors for Data-Efficient Dexterous Manipulation

TL;DR

The paper proposes a two-stage method that transfers control knowledge from simulation-based reinforcement learning into visual policies using a latent action decoder. It enables data-efficient learning from few visual demonstrations by training a visual encoder to predict latent task contexts that drive the decoder, without requiring privileged labels. The approach achieves superior performance with significantly fewer demonstrations than existing methods.

Why It Matters

This enables dexterous manipulation tasks to be learned from minimal visual data, reducing reliance on large-scale labeled datasets or extensive real-world training. It bridges the gap between simulation-acquired control policies and real-world visual perception, making advanced robotic manipulation more practical and scalable.

Key Findings

• The latent RL prior framework improves data efficiency by reusing pre-trained control knowledge from simulation through a flow-matching action decoder.
• Visual adaptation using action supervision alone outperforms methods that route predictions through a pretrained state encoder, improving both performance and training stability.
• The decoder's latent representations retain rich control information, such as contact forces and robustness to mass variations, indicating effective transfer of RL-acquired skills.