AI robot controllers raise average task success from 33% to 95%

Across 19 benchmark tasks, the authors report that allowing controlled departures from a pretrained policy outperformed stricter ways of keeping reinforcement learning close to demonstrations.

Top University

Charles Xu · Jacob Berg · Liyiming Ke · Sergey Levine

UC Berkeley · Physical Intelligence

Research Digest·4 November 2026·2 min read

Illustration: AI robot controllers raise average task success from 33% to 95%

In one line

Geometric Support Behavior Regularization lets an online reinforcement learning policy improve on a pretrained policy by penalizing distance from support, raising average success from 33% to 95%.

Before you read this

An idea this paper builds on

reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.256 papers →

Read it as

Trial-and-error robot learning is slow and costly. A common shortcut is to start from a pretrained policy, a set of behaviors learned from demonstrations, and refine it. Forcing imitation of that starting behavior blocks improvement; ignoring it wastes the prior.

GSBR, Geometric Support Behavior Regularization, is a penalty that guides this refinement. Actions inside the starting policy's known behavior set are unpenalized, so the robot can discard poor options. Actions outside are penalized smoothly by distance, so improvements remain possible when they are valuable enough.

Across a range of robot tasks, the average success rate rose from 33% for the starting behavior to 95%, beating all comparison methods. This makes trial-and-error fine-tuning of large demonstration-trained models practical for real robots.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

geometric support behavior regularization sample-based support approximation mode selection local refinement reinforcement learning behavioral cloning