Awaken-HI Human experience. Humanoid intelligence.

Scaling Humanoid Intelligence with Whole-Body Human Experience

Alternate view
0:00 0:00

Household tidying · continuous rollout

Highlight

  • Awaken-HI uses a shared action expert for humanoid loco-manipulation to predict original human motion together with robot actions.
  • Awaken-Human aligns 50,000 hours of first-person stereo recordings with full-body human motion across diverse activities and settings.
  • With more human pretraining data, action prediction error falls along an empirical power law and real-robot success improves. Awaken-HI generalizes to new object–destination pairings and objects unseen during robot post-training.

Model

Awaken-HI uses a shared action expert so that original human motion directly trains the network that predicts T800 actions. During pretraining, three branches learn to predict human body and hand motion, retargeted T800 actions, and camera-frame wrist poses using flow matching. They share a state encoder and a diffusion transformer that cross-attends to the image and instruction features from Qwen3-VL. On the robot, the T800 action branch uses the head-camera RGB image and robot state to generate a sequence of 40-dimensional whole-body targets for the low-level controller.

Awaken-HI architecture with ego pretraining on a towel-folding example and robot fine-tuning on a cart-picking example.

Scroll to explore →

Two-stage training of Awaken-HI. Pretraining on Awaken-Human (left) also supervises current and future scene depth and a description of the current subtask. Fine-tuning on T800 teleoperation data (right) updates Qwen3-VL, the shared action expert, and the T800 action branch using only the robot action loss.

Pretraining data

Awaken-Human pairs 50,000 hours of egocentric video with synchronized whole-body human motion. The curated dataset spans more than 100 skills and contains 17.02 million episodes, 5.41 billion pose frames, and 11.18 million valid language instructions. A PICO headset and five body-worn trackers record demonstrations as participants adapt how they walk, stand, and grasp to each setting. The annotation pipeline recovers 3D hand motion from video and adds subtask descriptions and stereo-derived depth.

Awaken-Human examples pair first-person footage with whole-body motion retargeted to T800 for action supervision.

Scaling human experience

We evaluate four nested pretraining subsets (5k, 12.5k, 25k, and 50k hours) on the same 100-hour held-out set. The scaling fits compare checkpoints at equivalent epochs. From the smallest subset to the full corpus, inference MSE on normalized T800 actions falls by about 11%, and validation flow-matching loss falls by 18.8%. Both metrics follow empirical power-law scaling with the amount of human pretraining data.

Inference MSE, validation loss, held-out scaling, and real-robot success rates at 10%, 25%, 50%, and 100% of the pretraining data.
At each scale, robot post-training starts from the checkpoint with the lowest inference MSE. With fixed robot data, mean success on three household tasks rises from 56.7% to 90.0%.

Real-world evaluation

Awaken-HI averages 93.75% success on individual T800 tasks, ahead of π₀.₅ and GR00T N1.7. Each method post-trains one multitask policy on the same 150 hours of teleoperation data and runs ten trials per task from matched starting configurations. Long-horizon trials succeed when the robot completes every stage in order without resets.

Industrial stocking task

The robot picks black material from the cart and stocks the middle shelf, then picks white material and stocks the upper shelf.

Household tidying task

The robot sorts a cup and toy car into separate bins, hangs a garment, then moves a basket out of the wardrobe.

Black Material Picking


Conclusion

Awaken-HI uses one action expert to predict both original human motion and retargeted robot actions. Awaken-Human aligns 50,000 hours of first-person stereo recordings with whole-body motion. With more pretraining data, action prediction error on held-out examples falls along an empirical power law, and success on downstream tasks improves. After robot post-training, one policy averages 93.75% success on eight individual loco-manipulation tasks. The same policy generalizes to objects unseen during robot post-training and to new pairings of familiar objects and destinations. For more details, see the paper.


Citation

@misc{engineai2026awakenhiscalinghumanoidintelligence,
      title={Awaken-HI: Scaling Humanoid Intelligence with Whole-Body Human Experience},
      author={EngineAI},
      year={2026},
      eprint={2611.26420},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2611.26420},
}