RoboMimic

## RoboMimic RoboMimic is a modular framework for robot learning from demonstration (LfD), developed by researchers at Stanford Vision and Learning Lab (SVL), NVIDIA, and the University of Texas at Austin as part of the ARISE Initiative. ### Key Paper - **Title**: "What Matters in Learning from Offline Human Demonstrations for Robot Manipulation" - **arXiv**: 2108.03298 (CoRL 2021) - **Authors**: Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, Roberto Martin-Martin ### Dataset Contents The robomimic v0.1 dataset collection includes: **Simulated Tasks (robosuite):** - **Lift**: Pick and lift a block - **Can**: Pick a can and place it in a bin - **Square**: Insert a square peg into a square hole - **Tool Hang**: Hang a wrench onto a peg - **Transport**: Move an object between two robot arms (dual-arm) **Dataset Variants:** - **Proficient-Human (PH)**: High-quality demonstrations from skilled operators - **Multi-Human (MH)**: Demonstrations from multiple human operators of varying skill levels - **Machine-Generated (MG)**: Synthetic demonstrations from trained RL/IL policies **Real-World Tasks:** - Real-world counterparts of Lift, Can, and Tool Hang tasks - Collected via RoboTurk teleoperation system ### Format Each dataset contains: - RGB images from multiple camera viewpoints - Robot proprioception (joint positions, velocities) - End-effector actions (position, orientation, gripper) - Task metadata and demo statistics ### License MIT License