Scaling Dexterous Manipulation

Panda-Allegro setup with Xela tactile sensors

We are developing scalable real-to-sim-to-real learning pipelines that combine behavior cloning from real-world demonstrations with reinforcement learning (RL). Demonstrations collected on real robots provide an effective warm start for RL, enabling curriculum-style training that extends behavior beyond the demonstrations. Our platform focuses on dexterous manipulation with multi-fingered Allegro hands equipped with Xela tactile sensors, allowing us to investigate how tactile feedback supports sim-to-real transfer. This integration of demonstrations, RL, and tactile sensing aims to accelerate policy learning while achieving robust performance in complex manipulation tasks.

Our goal is to enable real-world manipulation systems to learn complex skills without requiring large-scale real-world data collection.