Pieter Abbeel team speeds robot RL up to 1,000x with world model

The result comes from a group led by Pieter Abbeel, a professor at the University of California, Berkeley who is also affiliated with Amazon's Frontier AI & Robotics (FAR) organization. Abbeel is a well-known figure in robot machine learning; about a decade ago he paused his university teaching to work full-time as a researcher at OpenAI when that company was still doing robot reinforcement learning. He is also the mentor of Chelsea Finn, a co-founder of Physical Intelligence, and a former research collaborator of Sergey Levine.

The work sits in the classical lineage of world models in robot learning rather than the recent video-generation wave. WAM (world action model) and VAM (video action model) approaches, such as NVIDIA's DreamZero covered in the April 2026 issue of this magazine, are built mainly on diffusion-based video generation and are being applied to manipulation tasks. Abbeel's method instead directly predicts the state quantities of the robot and its environment as parameters.

The body notes that sim-to-real is currently used extensively for humanoid walking (locomotion) and full-body motion control such as dance. If the speed and success-rate gains hold, this could represent a significant step for that area, though the findings are presented as a university research result rather than a deployed product.