World-model RL trains Unitree Go1 to walk with 16 min of real data and 3 min on a GTX 3080
MarcJSchmidt · x · 2026-10-30
Developer MarcJSchmidt reports a major step in world models / physical AI: his model trains a Unitree Go1 quadruped to walk using only 16 minutes of recorded real-world data, under 4k policy updates, and 3 minutes of GPU time on a GTX 3080, with 100k total parameters.
The approach is MBRL via BPTT with a generic shared world model that works across arbitrary robots and can act as a foundation model, plus stable long-horizon dreams. It requires no simulator at all — no pre-training, no priors, no dense rewards, no privileged simulator data — and the author can now iterate on reward functions in minutes while watching the agent learn from scratch in real time. He says the bottleneck is now designing better reward functions.
More from Embodied
- Hugging Face co-founder Thom Wolf: give this robot an Oscar — Thom_Wolf · 2026-10-30
- Musk on Tesla vs Waymo robotaxis: 'It was never a real contest' — AccBalanced · 2026-10-30
- LA adds 11,000 aerospace jobs in two years as physical AI and defense tech boom — builditwithjoe · 2026-10-30
- Pokebot launches RoboAPI: a public API for near-real-time robot action prediction — CyberRobooo · 2026-10-30
- AGIBOT unveils 'Worken' metric: robots × autonomous hours × output efficiency — xmercury_one · 2026-10-30
- SCRUB-1: Fine-tuned π0.5 Robot Arm Lifts Surgical Instrument Handoff From 0% to 79% — StewartalsopIII · 2026-10-30