DexNex is a bimanual dexterous testbed with a rich immersive teleoperation interface. The teleoperator has no privileged information: they see only what the robot sees and feel only what the robot feels. This allows the robot to learn autonomous manipulation policies from teleoperation demonstration. Specifically, this project is exploring behavior cloning via Diffusion Policy.

The behavior cloning approach to generating autonomous manipulation policies will be augmented by reinforcement learning based on simulations from the DexNex Drake simulation project.

Diffusion policy learns an action-distribution score function.