ALOHA / ACT Demonstrations

ALOHA / ACT Demonstrations are human-teleoperated bimanual manipulation trajectories collected with the low-cost, open-source ALOHA hardware and used to train ACT (Action Chunking with Transformers) imitation-learning policies (Zhao et al., 'Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware', arXiv:2304.13705, RSS 2023). The paper reports 50 demonstrations recorded per fine-manipulation task (100 for Thread Velcro) across real-world tasks such as Slide Ziploc, Slot Battery, Open Cup, Thread Velcro, Prep Tape, and Put On Shoe, plus two simulated MuJoCo/dm_control tasks (Transfer Cube and Bimanual Insertion). Each episode takes 8-14 seconds to perform, corresponding to 400-700 timesteps at a 50Hz control frequency. Each timestep records 4 RGB images at 480x640 resolution (cam_high, cam_low, cam_left_wrist, cam_right_wrist), 14-dimensional joint positions/velocities/efforts for the two 6-DoF-plus-gripper arms, and a 14-dimensional absolute-joint-position action, all stored in a per-episode HDF5 file (schema documented in the repo's record_episodes.py). The demonstrations carry no per-step reward labels; they are behavior-cloning data whose downstream success is measured by per-task success rate on the robot. The teleoperation/recording code and the ACT algorithm are released under the MIT License, and scripted+human demonstration files for the two simulated tasks are distributed via a public Google Drive link from the ACT repo.