PyRUA-REAL: Robot Exploration and Adaptive Learning through RSI in the Real World

PyRUA-REAL. Robot Exploration and Adaptive Learning through RSI in the Real World. By Qiang Wang and nine others, from Peking University, EngineAI, and one other institution. Published on the fourth of November.

Coding agents, powered by models such as GPT-6-Astra, connect spatial reasoning with executable robot behaviour. An agent like that can write code that makes a robot move. The open question is what happens over time. How can such agents turn physical interaction into lasting improvements in what a robot is able to do?

The authors study embodied recursive self-improvement, or RSI, carried out through real-world exploration, code refinement and deployment. Their framework for it is PyRUA-REAL. In it, coding agents retain their interaction experience in two forms: as persistent memory, and as reusable executable skills. Both feed back into subsequent exploration and improvement, so each round starts from what the rounds before it left behind.

The empirical study places physical manipulation at the centre of the improvement process. It examines three things. How agents collaborate. What kinds of execution feedback support effective code refinement. And how task specifications and previously acquired skills influence robustness and transfer to new tasks. Each of those is a question about the loop itself rather than about any one task: who works with whom, what the robot's execution reports back to the agent, and what the agent brings to the next attempt. A further question concerns deployment. Once the improvement has happened, how efficient and how capable are the resulting programs when they run without an online agent involved in making decisions? That separates learning from use. The agent does its work during exploration and refinement, and what ships is the code it produced.

Across eight real-world tasks, PyRUA-REAL achieves an average success rate of ninety-one point one percent. A complementary evaluation uses forty-two tasks from the RoboDojo benchmark, and there its average success rate is eighty-four point four percent.

The study is offered as empirical guidance for turning physical exploration into reusable robot skills, skills that support continued improvement as well as robust, efficient deployment. The arc runs from an agent that explores and refines code on a real robot to programs that can be deployed without that agent in the loop.