PyRUA-REAL: Robot Exploration and Adaptive Learning through RSI in the Real World
TL;DR Two coding agents write, test and revise robot programs on a real dual-arm Franka, keeping what they learn as memory and reusable skills. Seven of eight real tasks end with programs that run without any language-model calls, and on 42 RoboDojo simulation tasks the frozen programs reach 84.4% success.
What they do
PyRUA-REAL studies whether coding agents can turn real-world trial and error into robot skills that last. Agents write manipulation programs for a dual-arm Franka, run them, diagnose failures and revise the code. What they learn is kept as memory and reusable skills, and the final program is frozen so it runs without language-model calls.
How it works
- Two agents. A researcher picks the approach, the experiment goals and the stopping criteria. An engineer implements the code, runs robot trials and reports results, findings, evidence and code back.
- Experience repository. It has three levels: facts about the robot and workspace, memory of attempted approaches and failure causes, and task programs plus reusable primitives.
- Hardware. Two Franka Research 3 arms with Robotiq-85 grippers, a head ZED Mini and wrist RealSense D405 cameras. Calibration and controllers are not given; the agents must build them. Programs may call SigLIP and SAM for perception.
Results
- Seven of the eight real tasks (tic-tac-toe, stacking, Hanoi, towel folding, Push T, sorting, storage) get agent-free programs. Time to first success ranges from 1.5 h to 45 h. Success across development attempts, pooled over code revisions, is 91.67-100% on these seven.
- Improvement continues after the first success: stacking cycle time falls from 285.3 to 156.0 s, and Hanoi full-cycle success rises from 60% to 100%.
- On 42 RoboDojo simulation tasks the frozen programs reach 84.42%, against reported 50.35% for PyRUA-Lean and 22.48% for GPT-6-Astra.
- Frozen programs are much faster: online coding takes 4.6-5.0x as long on tic-tac-toe and stacking, and neither online baseline finishes Hanoi within 120 min.
Limitations
Ethernet insertion still needs an online agent: 0/4 with the scripted program, 10/17 with online decisions. A frozen program cannot handle situations outside its encoded routines. Agents sometimes stalled on one method until a human suggested another. Most findings come from single development histories, and the RoboDojo baselines come from separate evaluations with different seeds.
Figures
Abstract
Coding agents powered by models such as GPT-6-Astra connect spatial reasoning with executable robot behavior, raising the question of how such agents can produce lasting improvements in robot capabilities from physical interaction. In this work, we study embodied recursive self-improvement (RSI) through real-world exploration, code refinement, and deployment using PyRUA-REAL, a framework in which coding agents retain interaction experience as persistent memory and reusable executable skills that inform subsequent exploration and improvement. Our empirical study places physical manipulation at the center of the improvement process, examining how agents collaborate, what kinds of execution feedback support effective code refinement, and how task specifications and previously acquired skills influence robustness and task transfer. We further investigate the efficiency and capability of the resulting programs when deployed without involving an online agent for decision-making. Across eight real-world tasks, PyRUA-REAL achieves an average success rate of 91.1%. In complementary evaluation on 42 tasks from the RoboDojo benchmark, it achieves an average success rate of 84.4%. This study provides empirical guidance for turning physical exploration into reusable robot skills that support continued improvement and robust, efficient deployment.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.