charlesxu0124/RLDG

This repository contains OpenVLA and Octo checkpoints fine-tuned on tele-op and RL generated datasets for the Connctor Insertion task in the paper RLDG: Robotic Generalist Policy Distillation via Reinforcement Learning .

Each model is trained on 45 episodes either collected by human tele-operation or rolling out converged RL policies. Each dataset contains 15 episodes for a VGA connector, 15 episodes for a USB-A connector, and 15 episodes for an ethernet connector. The models take 1 wrist camera and 1 language command as input, and outputs a 6D end-effector twist expressed in the wrist frame.