Work

LoopROS: A Deployable Embodied Recursive Self-Improvement Framework Based on Logical and Physical Constraints

Year: 2026
Type: preprint
Abstract: Embodied recursive self-improvement (RSI) endows robots with the capacity to progressively refine manipulation skills through real-world physical interaction, representing a cornerstone toward generalpurpose embodied intelligence. Yet existing embodied RSI frameworks largely remain simulationbound: they fail to realize closed-loop RSI when transferred onto uncalibrated real-world robots. This paper presents LoopROS, a deployable RSI framework built upon a constraints-as-policy principle for embodied robot. With In-Context Learning-Based Logical Constraints, the LoopROS parses real-world demonstrations, to guide segment selection, outcome assessment, and experience reuse. With Segmentwise Trajectory Tuning (STT)-based physical constraints, global trajectory editing error induced by accumulated disturbances during complex real-robot deployment can be greatly mitigated. The LoopROS achieves an 80.33% success rate and a mean score of 83.88, ranking No. 1 in RoboDojo benchmark. To the best of our knowledge, the LoopROS also provides one of the earliest RSI demonstrations on real robots, elucidating its great potential as a backbone embodied RSI.
Language: English
Open Access status: green
OpenAlex ID: W7234962212
Cites:
Cited by:
Related to: 10