EndoRSI improves endoscopic robot policies by learning from their physical failures

A code agent reviews the controller based on real tests under human authorization. The insertion success rate goes from 2 out of 21 to 17 out of 21 between generations.

EndoRSI organizes a search loop where an agent modifies the policy of an endoscopic robot based on physical experiment data. Execution remains isolated in a constrained environment and subject to human authorization.

Real-world trials reveal dangerous behavior that offline data did not allow to see, and then guide revisions. The fourth generation succeeds in seventeen out of twenty-one insertions, compared to two for the first. The final policy is evaluated on six ex vivo porcine specimens: this preclinical work does not describe a device authorized to be operated on patients.

Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.