Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents
Teaching robots to improve themselves by practicing in simulation before real tasks
Researchers developed RPG, a system that lets robots autonomously improve their performance without retraining their AI models. The system identifies what skills a robot can already do, creates practice tasks in simulation to diagnose failures, develops new reusable skills to fix those failures, and then tests improvements before deploying them—boosting success rates from 29% to 95% over 15 rounds of practice, and achieving 100% success on physical robot trials.
Robots currently require constant human effort to fix failures and add new capabilities. RPG eliminates this bottleneck by letting robots diagnose and fix their own problems through simulation practice, then carry those improvements to physical tasks. This could dramatically reduce the engineering overhead needed to deploy robots in real-world settings like warehouses, manufacturing, and other manipulation-heavy environments.