← News

Agents & MCPSources: arXiv · 18 Nov 2026

AdaRSI lets a coding agent retrain its robot model on demand

An arXiv paper describes a framework where an agent diagnoses failures, collects demonstrations, and trains a specialist model to assist its robot executor.

In plain words

Robots powered by vision-language-action models can fail at long tasks either because of how the model is prompted or because of its underlying skills. AdaRSI is a framework where a coding agent fixes its own tools and instructions first, then gathers robot demonstrations and trains a helper model to handle steps that stay unreliable. The authors report success-rate gains on simulated benchmarks and a real robot.

What was announced

The paper, an arXiv abstract by the AdaRSI authors, frames failures in long-horizon embodied tasks as coming from two places: how a vision-language-action model is called, and what the model can actually execute. AdaRSI works recursively. A coding agent first refines the tools and procedures it uses, then targets sub-steps that remain unreliable after testing alternative instructions. The agent collects demonstrations and checks their value through temporary fine-tuning probes before accepting them.

Accepted data is pooled across tasks to train a specialist VLA that runs alongside the original executor, and each development round is summarized into shared experience that guides later diagnosis and data collection. The authors report that on unseen composite tasks in RoboCasa365, success rises from 4.6% to 13.3%, and on perturbed LIBERO-Pro 2.0 tasks from 32.9% to 49.1%. On RoboCasa365's Composite-Unseen suite, they say it beats adapting every manipulation sub-step while using 34% of the GPU time and 36% of the demonstrations. On a real robot, the agent autonomously collected demonstrations and adapted the executor to recover fallen bottles.

What it does not show

  • The evidence is the paper's abstract only; the reported benchmarks and real-robot results have not been independently verified, and the excerpt does not cover the full paper's methods or failure cases.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.