RPG revises robot skills through simulated practice
Researchers improve robot skills and prompts in simulation
An October 1 2026 preprint introduces Reconstruct, Practice, Go Real, or RPG, which revises robot skills and prompts through simulated practice without updating model weights.
Failure diagnosis uses privileged simulator state and source videos. Candidate and merged revisions must improve average performance across tasks.
The authors report 209 successes in 220 episodes after 15 rounds on 22 simulation tasks. Evaluation used held out initializations from the same task distributions. The largest improvement round changed prompting alone.
What the physical trials show
The frozen system completed all 30 physical trials across three tasks after calibration and hardware adaptation. Trials allowed autonomous retries but no human assistance, capped at 20 minutes or 30 model calls. Physical towel folding used different criteria from simulation.
The project page says code is coming soon. These author reported results come from one improvement run. Independent reproduction and reliability on unfamiliar tasks remain unestablished.
Illustrative archival photograph of the NIST Dexterous Manipulation Testbed from July 2013 by Falco/NIST, public domain in the United States. NIST reuse information. Resized and converted to WebP.
