Awaken-HI: Scaling Humanoid Intelligence with Whole-Body Human Experience
TL;DR EngineAI records 50,000 hours of egocentric stereo video with synchronized whole-body human motion and pretrains a humanoid VLA to predict both the original human motion and the retargeted robot actions. After post-training on 150 hours of teleoperation it averages 93.75% success on eight real T800 loco-manipulation tasks, against 57.5% for π0.5 and 58.75% for GR00T N1.7.
What they do
Whole-body teleoperation is too costly to cover many humanoid tasks, and egocentric human video gives no whole-body motion targets. The authors record Awaken-Human: about 50,000 hours of egocentric stereo video with synchronized whole-body motion, covering more than 100 skills in homes, workplaces, shops and storage areas. They pretrain a VLA, Awaken-HI, on it, then post-train it on the EngineAI T800 humanoid.
How it works
- Capture and annotation. A PICO headset and five body trackers record stereo video and skeletons. An in-house 3D hand model replaces the PICO hand tracking, stereo matching supplies depth, and a VLM proposes subtask labels that people correct.
- Retargeting. A T800-specific pipeline built on GMR cuts mean end-effector position error from 16.4 cm to 2.4 cm.
- Shared action expert. A Qwen3-VL backbone feeds one flow-matching diffusion transformer with three branches: retargeted 40-D robot actions, the original 161-D human motion and camera-frame wrist poses. Extra heads predict current and future depth and the current subtask.
- Post-training on about 150 hours of T800 teleoperation, using only the robot-action loss.
Results
- On the real T800 (10 trials per task), Awaken-HI averages 93.75% over eight tasks, against 57.5% for π0.5 and 58.75% for GR00T N1.7. On the long-horizon industrial and household tasks it reaches 90% and 80%, while neither baseline exceeds 30%.
- Growing pretraining data from 5,000 to 50,000 hours lowers held-out action MSE by about 11% (power-law exponent 0.051). Mean success on three household tasks rises from 56.7% to 90.0%.
- In the 5,000-hour ablation, adding human-motion supervision lifts household success from 13.3% to 43.3%, and all auxiliary targets together reach 53.3%.
- New object-destination pairings succeed 80-100% of the time, against 0-40% for the baselines.
Limitations
The paper has no limitations section. All results come from one humanoid with ten trials per condition, and the scaling and ablation robot tests cover only three household tasks. Transfer to unseen objects is partial: 50% completion for tissues and 30% for a medicine bottle, though both baselines score 0%.
Videos
Figures




Abstract
A central goal of humanoid robotics is for robots to perform the everyday work people do in homes and workplaces. Existing egocentric video datasets cover many of these activities but provide little synchronized whole-body motion data for learning loco-manipulation at scale. We build Awaken-Human, which pairs approximately 50,000 hours of egocentric stereo video with synchronized whole-body human motion and covers more than 100 skills. Building on this, we also introduce Awaken-HI (Humanoid Intelligence), a vision-language-action model for humanoid loco-manipulation that jointly predicts original human motion and retargeted robot actions through a shared action expert. We show that loco-manipulation skills transfer more effectively from people to humanoids when human trajectories directly supervise the shared action expert. Increasing human pretraining data reduces held-out robot action prediction error according to an empirical scaling law and raises the mean success rate on downstream real-robot tasks. After full-corpus pretraining and T800 adaptation, Awaken-HI achieves a mean success rate of 93.75% across eight individual tasks and outperforms GR00T N1.7 and $π_{0.5}$ in both individual and long-horizon evaluations. The same policy generalizes to objects unseen during robot post-training and to new pairings of familiar objects and destinations.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
