USC lab open-sources Psi-Zero humanoid foundation model trained on just 800 hours of human video
yuewang314 · x · 2026-09-16
USC PSI Lab (with NVIDIA and WorldEngine) released Ψ₀ (Psi-Zero), an open foundation model for humanoid loco-manipulation, accepted to RSS 2026, with code, weights and data public.
Key points:
- The team argues co-training on mixed human/humanoid data is suboptimal due to fundamental kinematic disparities.
- Staged paradigm: autoregressive VLM pre-training on large-scale egocentric human manipulation videos, then post-training a flow-based action expert on high-quality real humanoid trajectories for precise joint control.
- Data recipe insight: quality egocentric human video pre-training beats scaling with noisy internet clips or heterogeneous cross-embodiment data.
- Achieves best real-world performance with only 800 hours of human video data.
The author is also running an experiment paying the robot a crypto wage, starting with wiping his desk, to document "a week of paid work for a humanoid."
Related event: USC Open-Sources Humanoid Model Psi-Zero, First Crypto-Paid Task Completed(2 posts)→
More from Embodied
- Tesla logs 460,000 km testing FSD Supervised across Spain over two years — yunta_tsai · 2026-09-16
- VLA-Replica: low-cost real-world robot benchmark, MolmoAct2 beats π0.5 by 10% OOD — YuXiang_IRVL · 2026-09-16
- SSAD 2026: dedicated reward models for crashes and sim-driven failure analysis in driving — abursuc · 2026-09-16
- 90 km/h on mountain roads: Swaayatt trains RL agents for L5 driving — sanjeevs_iitr · 2026-09-16
- Humanoid robot Izumi shows off leg-day moves in new demo video — poppy-slashpot · 2026-09-16
- Moving Parts unveils Bingo, an AI companion toy that makes expression the interface — round · 2026-09-16