robomimic

A learning-from-demonstration framework from Stanford and UT Austin researchers, with standard datasets and offline learning algorithms.

robomimic is a learning-from-demonstration framework developed under the ARISE initiative by researchers at Stanford and UT Austin, with its paper published at CoRL 2021. It provides data for five manipulation tasks in the robosuite simulator — Lift, Can, Square, Transport, and Tool Hang — in three variants: PH is 200 demonstrations from one skilled operator, MH is 300 demonstrations from 6 operators of varying skill, and MG is data generated by a reinforcement-learning agent. The framework includes built-in behavioral cloning, BC-RNN (behavioral cloning that can look at history), and several offline reinforcement-learning algorithms, with diffusion policies added in later versions. The paper found that models which look at history perform better, and that data quality matters a great deal; this dataset later became a standard simulation benchmark for imitation-learning papers.

ExampleThe Diffusion Policy paper compared success rates against BC-RNN and other methods on robomimic's Lift, Can, Square, Transport, and Tool Hang tasks.

Related

robosuite · Behavior Cloning · Offline Reinforcement Learning · Demonstration Data · Diffusion Policy · RoboTurk

Sources

robomimic 官网 (Chinese)
robomimic v0.1 数据集说明 (Chinese)
What Matters in Learning from Offline Human Demonstrations for Robot Manipulation (arXiv 2108.03298)