ARISE-Initiative/robomimic

What it solves

Robot learning often suffers from a lack of standardized datasets and reproducible benchmarks, making it difficult for researchers to develop and compare new algorithms for robot manipulation.

How it works

robomimic provides a comprehensive framework for learning from demonstration. It combines a collection of standardized demonstration datasets (both simulated and real-world) with a suite of offline learning algorithms. The system supports various observation modalities, including low-dimensional data and visuomotor policies, and allows for flexible network architectures and hyperparameter tuning.

Who it’s for

Researchers and practitioners in robotics and AI who want to benchmark robot manipulation tasks and develop new robot learning algorithms.

Highlights

• Diverse Learning Algorithms: Supports Imitation Learning (BC, BC-RNN, HBC) and Offline RL (BCQ, CQL, IRIS, TD3-BC), as well as newer additions like Diffusion Policy and IQL.
• Standardized Datasets: Includes simulated and real-world tasks across multiple environments and robots.
• Modular Design: Features modular observation encoders and support for external datasets.
• Developer Tools: Includes tools for dataset visualization, hyperparameter sweeps, and Docker support for easy environment setup.