Robots Learn Without Retraining

🔥 What happened Berkeley researchers dropped RPG, a framework that teaches robots new manipulation skills without touching model weights. It reconstructs practice tasks in simulation, diagnoses failures, and rewrites its own skill library and system prompt. 💡 Why it matters Success rate on 22 held-out tasks jumps from 28.6% to 95.0% after 15 practice rounds – crushing GPT-6-powered baselines stuck at 60%. All 30 physical trials passed. Robot skills now scale without expensive finetuning. ⚡ Our take If you're still training weights, you're already behind. The future belongs to systems that debug themselves – just like good engineers do.