Researchers developed a robot system that practices related tasks in simulation, then revises shared manipulation skills without changing model weights. In tests on 22 simulated tasks, success reached 95% after 15 practice rounds; after a common hardware-calibration procedure, the system succeeded in all 30 trials across three physical tasks.
Reconstruct, Practice, Go Real, or RPG, selects capabilities from ABC, an offline dataset spanning 193 manipulation tasks, then builds related practice scenes in simulation rather than replaying original demonstrations. The evaluation suite combined seven CaP-X tasks with 15 ABC-derived or analogous tasks.
For each practice task, a Runtime Agent tries the job using the observations available to a deployed robot. A second agent repeats it from the same starting conditions with access to simulator state. A video analyzer compares their attempts and available dataset videos to locate failures; an implementer then updates a shared library of robot skills or the system prompt that tells the agent how to use them.
The system tests each proposed change across the task suite before keeping it. A revision must improve average success without lowering any single task by more than 20 percentage points, and merged changes are tested again. The model weights, task definitions and success checks stay fixed, so improvement comes from revising reusable procedures and instructions rather than retraining the model.
In one 15-round run, the team evaluated each saved version on the same ten held-out starting setups for each of the 22 tasks, or 220 episodes per version. Success rose from 28.6% after the first round to 95.0% after round 15, equal to 209 completed episodes out of 220.
On the same 22 tasks and held-out starting setups, ASPIRE scored 75.5% and CaP-Agent0 running GPT-6 Astra Pro scored 60.0%. RPG and ASPIRE were limited to 30 model turns per episode, while CaP-Agent0 could use up to 50, so the comparison uses the same tests but unequal interaction budgets.
After a common hardware-calibration procedure, RPG succeeded in all ten trials for each of three tasks: placing a ball in a drawer and closing it, folding a towel, and transferring a bowl between robot arms. The YAM robot could retry or recover on its own, but received no human assistance, manual reset or system revision during a trial; each trial ended after 30 calls to the vision-language model or 20 minutes, whichever came first. Five further physical tasks were shown only as qualitative examples.
The study is a new arXiv preprint. Its authors cite API limits on simulation rollouts and the difficulty of combining overlapping skill changes as barriers to scaling the practice process. The physical towel-folding success criterion also differed from the one used in simulation.
Sources
- Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents, arXiv:2610.02204v1, submitted October 1, 2026.
- Figure 1 workflow diagram from the authors’ paper, licensed under CC BY 4.0.