Adaptive self-improving system boosts robot success on unseen tasks
AdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents
Computation and LanguageRobotics
Summary
Robots often struggle to complete long tasks they haven't seen before. The authors created AdaRSI, a system that lets a robot improve itself only when needed, by figuring out exactly where it fails and fixing just those parts. This method saved a lot of computing time and helped the robot get better at new tasks without needing human help. Their approach improved success rates significantly in tests on many robot tasks.
What this means in practice
- •For robotics engineers: Improve robotic agents' task success by selectively retraining only failing task parts, reducing training cost and increasing efficiency on complex tasks.
- •For autonomous system developers: Build adaptive autonomous systems that identify and correct specific failures during execution without exhaustive retraining.
Authors
Peidong Wang, Ying Chang, Zhiming Ma, Xiaocui Yang, Shi Feng, Daqi Liu, Yuzhe Yang, Xu Cao, Chuang Wang, Jian Su, Daling Wang
Abstract
Pretrained vision-language-action models (VLAs) often fail midway through unseen long-horizon tasks. Recursive self-improvement (RSI) offers a remedy, but embodied RSI methods either refine the harness around a frozen executor and so cannot add a missing motion, or train the executor on persistent failures without asking whether a failure lies in how the executor is called or in what it can do. We ask when a self-improving embodied system should change the weights of its executor, and propose AdaRSI, an on-demand RSI framework for embodied agents. A coding agent first improves its skills and harness, then attributes persistent failures by retrying failing sub-steps from saved states, and fine-tunes only the sub-steps that remain unreliable, with demonstrations it collects itself and experience carried over from earlier tasks. On all 50 target tasks of RoboCasa365, with GR00T N1.5 as the executor and no human demonstrations, AdaRSI improves the VLA on all three suites and raises its success on unseen composite tasks from 4.6% to 13.8%. Compared with fine-tuning every sub-step, it reaches slightly higher success while using only 28% of the GPU time and 22% of the self-collected demonstrations, and fine-tuning without a harness trains more yet gains far less.