ExRL adapts long-chunk robot policies by learning execution length

Robots trained to predict long action chunks have been getting steadily better at manipulation. The hard part has been adapting them to a new task without burning days of physical interaction. A preprint from researchers at UESTC, Tongji University and The Hong Kong Polytechnic University argues the lever is not the chunk itself but how long the robot keeps running it before replanning.

Their method, called Execution at the Right Length (ExRL), leaves the base policy alone and trains a lightweight adapter that picks an execution length for each predicted chunk. The paper puts the pitch flatly: "Our key insight is to shift adaptation from optimizing actions to optimizing how long to execute each predicted action chunk." The authors add that this "keeps the adaptation space low-dimensional while its expressivity grows with the chunk size."

On real hardware, the paper reports approaching perfect success on cube stacking "with about ten minutes of physical interaction" and reaching "nearly 90% on the long-horizon clothing folding task within 85 minutes." On power insertion, the ExRL-Q variant hits 22 of 25 trials against zero of 25 for a fixed 50-step execution baseline. In simulation on LIBERO, the authors lift π₀.₅ "from nearly 0% to over 90% success within 200,000 environment steps."

The adapter is policy-agnostic across the paper's test matrix: Diffusion Policy, ACT, π₀, π₀.₅, π₀-FAST and Cosmos Policy. Caveats are inline. Real-world runs keep the base policy frozen and do not yet incorporate online behavior cloning; the authors concede "several tasks still fall short of 100% success." Multi-task adaptation and generalization to unseen tasks are listed as untested, as is compatibility with asynchronous inference such as real-time chunking.

The work lands amid a run of Chinese-lab real-robot policy results on our tracker, including SJTU and Noematrix's SWAM a day earlier.