robomimic
Stanford Vision and Learning Lab / ARISE Initiative · 2021 · datasets.bot · datasets.bot page
One-liner. Stanford imitation-learning benchmark: human + machine-generated Franka Panda demos across 5 sim and 3 real manipulation tasks.
Setup
- Datasets / benchmarks: robomimic is a framework and dataset suite for robot learning from demonstration, released with the CoRL 2021 study 'What Matters in Learning from Offline Human Demonstrations for Robot Manipulation' (Mandlekar et al.). It provides teleoperated human and machine-generated demonstrations on five simulated robosuite/MuJoCo tasks (Lift, Can, Square, Transport, Tool Hang) using the Franka Panda robot (Transport is dual-arm), plus three real-world Franka Panda tasks. Data is distributed as HDF5 with low-dim, image, and raw variants, and includes RGB camera views, proprioception, end-effector pose, and object state. License: MIT. Download: https://huggingface.co/datasets/amandlek/robomimic.
- Hardware / simulator: Embodiment: franka_panda. Environment: lab, simulation, tabletop. Realness: both.
Schema
episodes -> steps -> {observation: {agentview_image, robot0_eye_in_hand_image, robot0_eef_pos, robot0_eef_quat, robot0_gripper_qpos, joint states, object}, action, reward, done}
Links