Robotics / news brief
EngineAI trains Awaken-HI on 50,000 hours of human motion
- Event
- Published

EngineAI introduced Awaken-HI, a vision-language-action model pretrained on a claimed 50,000 hours of synchronized first-person video and whole-body motion, then post-trained on about 150 hours of T800 teleoperation. The authors report 93.75% mean success across eight real-robot tasks, with ten trials per task. The dataset is public, but inference code, model weights and the simulation platform are still promised rather than released.
Why it matters
Large human-motion corpora could reduce robot-specific data needs, but the headline gains remain an author-run preprint result that outside teams cannot yet reproduce.
Sources
- Awaken-HI: Scaling Humanoid Intelligence with Whole-Body Human Experience (external site)EngineAI · primary source ·
- EngineAI introduces Awaken-HI for humanoid training (external site)Data Phoenix · coverage ·