Why Anthropomorphic Robots? Human Video Data Helps a Humanoid Beat Nvidia GR00T

At Nvidia's GTC Taipei on June 1, Jensen Huang announced the Isaac GR00T humanoid reference platform, a full‑size robot with dexterous hands, perception, and onboard compute, open to researchers worldwide.

Nvidia framed the need for a unified, human‑like body as the carrier for general physical intelligence, positioning the robot body as core AI infrastructure.

Deepwise (深度机智) released the paper Human-as-Humanoid on June 30, demonstrating that a self‑developed humanoid, PrimeU, can learn complex real‑world tasks without any task‑specific robot demonstration data, using only action supervision derived from human videos.

The authors argue that the biggest bottleneck for embodied intelligence is data. Traditional tele‑operation data collection is slow, costly, and limited in diversity. To overcome this, Deepwise built a 4,000 m² data‑capture factory in Shanghai and multiple collection centers, emphasizing the need for massive "observation‑action" pairs.

By aligning the robot’s physical dimensions with the ANSUR II human body database (50th‑percentile male), PrimeU minimizes transfer error: shoulder‑width ratio 0.97, arm‑reach ratio 1.02, palm‑length ratio 1.00. The platform features two 7‑DOF arms, two 20‑DOF dexterous hands, a 3‑DOF neck, and a 3‑DOF waist, totaling 60 DOF.

Data capture abandons inertial motion‑capture suits and relies solely on cameras: a head‑mounted camera provides first‑person view, while an external RGB camera supplies a third‑person view for robust hand and arm reconstruction. Experiments show the pure‑vision pipeline yields more stable key‑point recovery than wearable systems, especially in close‑hand tasks.

Processing runs at ~20 fps: third‑person video is used to track human joints, which are then mapped to PrimeU’s joint space via a staged inverse‑kinematics solver, producing 60‑DOF action chunks directly usable for robot training. This pipeline achieves a 4.8–7.2× increase in data‑collection throughput compared with tele‑operation.

To exploit the generated action labels, Deepwise developed PhysDex , a high‑DOF VLA policy that predicts joint‑space action chunks. PhysDex incorporates a dual‑space hierarchical kinematic constraint (DS‑HKC) that adds geometric supervision on wrist pose and fingertip positions via differentiable forward kinematics, eliminating the need for extra annotations.

Ablation studies confirm that DS‑HKC reduces training loss and improves geometric accuracy, enabling the learned policy to execute reliably on real hardware.

For evaluation, the team pre‑trained on 1,500 hours of human demonstrations, converted to 60‑DOF labels, and tested on seven dual‑hand tasks (ring‑placement, Rubik’s‑cube packaging, cup‑stacking, pouring, temperature‑gun measurement, light‑bulb screwing, bottle‑cap loosening). The latter three incorporated a small amount of real robot data as anchors.

PhysDex outperformed Nvidia’s GR00T N1.7 on all seven tasks, with especially large margins on the pure‑human‑demonstration tasks. The normalized mean error of the reconstructed human actions was 0.008, and the end‑effector error after forward kinematics was 5.34 mm, comparable to the 4.09 mm error from models trained on robot data.

The results demonstrate that, with a suitably designed body and a vision‑only data pipeline, human video experience can be directly transformed into executable robot training signals, achieving zero‑sample task performance and establishing a full‑stack technical loop from data to model to deployment.

Looking forward, Deepwise plans to address occlusion and motion‑blur challenges in first‑person capture and to incorporate synthetic interaction videos, aiming to further accelerate the data‑flywheel for embodied AI.