Embodied AgentsExpert
Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents
This preprint introduces RPG, a framework that uses offline demonstration videos, simulator practice, and privileged-state rollouts to autonomously improve reusable robot skills and prompts without updating model weights. On held-out initializations of 22 manipulation tasks, RPG reaches 95.0% success after 15 practice rounds, beating ASPIRE at 75.5% and CaP-Agent0 with GPT-6 Astra Pro at 60.0%, and succeeds in all 30 physical trials across three tasks. The result matters because it shows shared skill code and a system prompt can be a practical target for autonomous robot capability improvement.