Awaken-HI: Scaling Humanoid Intelligence with Whole-Body Human Experience

EngineAIarXiv 2026 · arXiv

TL;DR

Awaken-HI scales humanoid intelligence by pretraining on Awaken-Human, a massive new dataset of 50,000 hours of paired egocentric video and synchronized whole-body human motion. By jointly predicting original human trajectories and retargeted robot actions through a shared "action expert," the model achieves a 93.75% success rate on real-world tasks and demonstrates that robot performance follows a clear scaling law relative to human pretraining data.

What Problem Does This Paper Address?

For humanoid robots to be useful in homes and workplaces, they must master loco-manipulation: the ability to coordinate walking and balancing with complex hand movements. Traditionally, robots learn these skills through teleoperation, but collecting high-quality whole-body teleoperation data is expensive and difficult to scale.

While the internet is full of human videos, most egocentric datasets lack the synchronized, high-fidelity whole-body motion data required to train a robot's "motor cortex." Furthermore, simply "retargeting" human motion to a robot's proportions often introduces errors and distortions. The authors address this by creating a massive-scale dataset and a model architecture that learns directly from the "natural" human motion prior to bridge the morphological gap between humans and robots.

Methods

The authors introduce Awaken-HI, a Vision-Language-Action (VLA) model designed for two-stage training:

  1. Egocentric Pretraining: The model is pretrained on the 50,000-hour Awaken-Human dataset. It uses a Qwen3-VL backbone to process visual and language inputs. The core innovation is a Shared Action Expert (based on a Diffusion Transformer) that jointly predicts three things: original human motion, retargeted robot actions, and wrist poses.
  2. Robot Post-training: The model is fine-tuned on approximately 150 hours of real-robot teleoperation data from the EngineAI T800 humanoid.
The two-stage training pipeline of Awaken-HI

To enhance perception, the model also performs auxiliary tasks during pretraining, such as predicting current and future scene depth and describing current subtasks.

Datasets

The primary contribution is the Awaken-Human dataset, which dwarfs previous efforts in this space:

  • Scale: 50,000 hours of egocentric stereo video (17.02M episodes).
  • Data Quality: Pairs video with synchronized whole-body motion (recorded via PICO headsets and wearable trackers) and 3D hand reconstructions.
  • Diversity: Covers 100+ skills across homes, workplaces, and retail spaces.
Object and skill coverage in the Awaken-Human dataset

The paper compares this to existing datasets like Ego4D (3,670h) and Nymeria (300h), highlighting that Awaken-Human is the first to provide whole-body motion at this scale.

Key Insights and Ideas

  • The Power of Human Priors: The authors found that loco-manipulation skills transfer more effectively when the model is supervised by original human trajectories alongside retargeted robot actions. This helps the model learn "natural" motion patterns that retargeting alone might distort.
  • Empirical Scaling Laws: One of the most significant findings is that robot performance follows a power law. As the amount of human pretraining data increases from 5,000 to 50,000 hours, the robot's action prediction error consistently drops, and its real-world success rate climbs.
  • Depth as a Visual Anchor: Auxiliary depth supervision is crucial. It helps the model's visual backbone "see" the relative depths of arms and surfaces more clearly, which is vital for precise manipulation.
  • Compositional Generalization: Because it learns general motion and task priors, the model can follow new instructions that pair familiar objects with new destinations (e.g., "put the black material on the upper shelf" when it was only trained to put it on the middle shelf).

Key Results

Awaken-HI was evaluated on the EngineAI T800 humanoid across eight complex tasks, including garment hanging and laundry basket relocation.

  • Superior Success Rates: Awaken-HI achieved a 93.75% mean success rate, significantly outperforming baselines like GR00T N1.7 (58.75%) and $\pi_{0.5}$ (57.5%).
  • Scaling Impact: On a subset of household tasks, the success rate rose from 56.7% (using 5k hours of data) to 90.0% (using the full 50k hours).
  • Zero-Shot Generalization: The model successfully manipulated objects it never saw during robot post-training, such as tissues (50% success) and medicine bottles (30% success).
Scaling results and generalization performance

Limitations and Future Work

While the results are impressive, the authors note some areas for growth:

  • Morphological Gap: Despite the shared expert, the physical differences between humans and robots still pose challenges for perfect motion transfer.
  • Sensor Requirements: The human data collection still relies on wearable sensors; moving toward learning from "in-the-wild" video without specialized trackers remains a major frontier.
  • Long-Horizon Complexity: While the model handles chained tasks well, truly open-ended, multi-hour autonomy in unstructured environments is the next big step.
The eight real-robot evaluation tasks

Summary updated

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.
BibTeX
@misc{engineai2026awaken,
  author = {EngineAI},
  title = {{Awaken-HI: Scaling Humanoid Intelligence with Whole-Body Human Experience}},
  year = {2026},
  eprint = {2611.26420},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26420},
}
Figure