Driving · 2026-10-15
AdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents
Peidong Wang, Ying Chang, Zhiming Ma, Xiaocui Yang, Shi Feng, Daqi Liu, Yuzhe Yang, Xu Cao, Chuang Wang, Jian Su, Daling Wang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
AdaRSI is an on-demand recursive self-improvement framework for embodied agents. A coding agent first refines a harness (skills/tools/procedure) using execution feedback, then attributes persistent failures to either harness calling or executor limitations by retrying sub-steps from saved states. It fine-tunes only sub-steps that remain unreliable, using self-collected simulation demonstrations and carries experience across tasks. On RoboCasa365 (50 tasks, no human demos), unseen composite success rises from 4.6% to 13.8%.
Why it matters
The authors target a key ambiguity in embodied recursive self-improvement: whether a persistent failure is due to how a vision-language-action executor is called (harness) or due to missing motion capabilities in the executor (weights). AdaRSI changes weights only when failures persist after harness refinement and validation, aiming to improve long-horizon task success with less training cost.
Method
- Refine the harness first: the agent writes/strengthens tools and the procedure that calls executor policies, using held-out validation feedback after each round.
- Attribute persistent failures: retry failing sub-steps from saved simulator states under different instructions/approaches to separate poor calling from executor-ineligible motions.
- Fine-tune on demand: for sub-steps that remain unreliable, collect self-collected simulation demonstrations (rejection sampling or scripted expert with privileged labels), fine-tune sub-step policies, validate, and adopt them only if they help.
Limitation
All experiments run in simulation, where privileged state makes it easy to generate and label demonstrations. On a real robot, privileged state would need to be replaced by success detectors, scripted resets, or occasional human help, and the cost of a weight change would rise accordingly.
Abstract (from arXiv)
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.
Related papers
- EmbodiedEvo: Self-improving Coding Agents for Physical Intelligence
- Toward Recursive Self-Improvement in Physical AI
- EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution
- SEAR: Self-Evolving Agentic Robotics
- UniWALM: Unified World Action Language Model for Embodied Robot Control
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-14 · Indexed by Pub-AI: 2026-10-15 UTC
AdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents
Preprint posted: 2026-10-14 · Original record ↗
- arxiv
- 2610.20474
Explicit source versions (1)
- v1 · 2026-10-14
- Preprint available
Source date: 2026-10-14 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.