AdaRSI preprint reports lower GPU use through selective robot fine-tuning

Researchers report that AdaRSI, a framework for improving embodied agents, used 28% of the GPU time and 22% of the self-collected demonstrations needed to fine-tune every task sub-step, while achieving slightly higher success. The preprint evaluates all 50 target tasks of RoboCasa365 with GR00T N1.5 as its action executor and no human demonstrations. A coding agent first adjusts its skills and supporting software, then retries failed sub-steps from saved states and trains only those that remain unreliable. Success on unseen composite tasks rose from 4.6% to 13.8%, leaving most attempts unsuccessful despite the improvement.