Berkeley RAIL

Deep publication record spanning learning-based control, dexterity and generalization.

No.Paper · what it does and showsIndustry contribution

DexterityGen: Foundation Controller for Unprecedented Dexterity2025 ↗

Pretrains large-scale dexterous motion primitives with sim-to-real reinforcement learning, distills them into a foundation controller, and uses human teleoperation only as a coarse prompt at run time, improving hold stability 10-100x and enabling pen, syringe and screwdriver use.

Moves tool use and long-duration in-hand manipulation within reach of robot vendors targeting assembly, lab and service work that needs human-grade hands.

BridgeData V2: A Dataset for Robot Learning at Scale2023 ↗

An open dataset of roughly 60,000 multi-task, multi-environment manipulation trajectories spanning kitchens, toys, sinks and diverse objects and scenes.

Became one of the most-used open manipulation datasets for pretraining policies and for benchmarking generalization across environments.