A new robotics framework teaches itself to do chores better by rehearsing in a simulator, no model retraining required.
The system, called Reconstruct, Practice, Go Real (RPG), starts with an offline dataset of robot manipulation attempts and builds matching practice tasks in simulation. During practice rounds, it uses execution feedback, privileged simulator state, and recorded videos to figure out why a task failed, then writes new reusable skills, fixes old ones, and edits the instructions, known as the system prompt, that guide a multimodal LLM controlling the robot. Each candidate change gets tested across tasks before it is kept. Across 22 held-out manipulation tasks, task success climbed from 28.6 percent after the first practice round to 95.0 percent after 15 rounds, beating a baseline called ASPIRE (75.5 percent) and a GPT-6 Astra Pro-powered agent called CaP-Agent0 (60.0 percent). After a calibration step for the real hardware, the frozen system went 30-for-30 on physical trials across three tasks.
That's a direct hit on the actual bottleneck in robotics, not better motors, but the human hours spent hand-coding skills, rewards, and perception pipelines for every new task. By letting the system diagnose its own failures and rewrite its own playbook in simulation, RPG swaps months of engineer time for compute cycles, without touching the underlying model's weights.
Thirty physical trials is a thin slice of the real world, and held-out initializations of known tasks is not the same as a robot improvising in someone's actual kitchen. The simulation-to-reality gap has eaten plenty of promising robotics demos before this one.