RAIN

RAIN learns robotic interactions conditioned on target regions and action types. It uses visual observations and robot state to predict actions and determine when a subtask is complete.

checkpoint.pt contains the LIBERO-trained action policy and Transition Head. The released checkpoint is for LIBERO simulation and uses third-person and wrist camera observations. Keep the accompanying config.json in the same directory.

See the RAIN code for installation and loading, and RAIN-LIBERO for training data. RGB inference requires external DINOv2 weights. CLIP action features are already stored in the checkpoint.