Research

HuMI: Robot-Free Whole-Body Skills on Unitree G1

Robb Harlan 5 min read

Teleoperating a humanoid for whole-body demos is slow and brittle. HuMI (Humanoid Manipulation Interface) from Tsinghua University, Shanghai Qi Zhi Institute, Spirit.AI, and Shanghai Jiao Tong University collects robot-free whole-body demonstrations with hardware that fits in a backpack, then trains a hierarchical stack that runs on a Unitree G1.

Paper: arXiv:2602.06643. Project: humanoid-manipulation-interface.github.io.

Portable capture, full-body intent

HuMI extends UMI-style handheld grippers (wrist GoPros) with HTC Vive Ultimate trackers on grippers, waist, and feet — five operational frames including the pelvis base. Operators see a real-time IK preview of the G1 so they can adjust motions that would be unreachable or self-colliding if you naively scaled adult human trajectories onto a ~130 cm robot.

A high-level Diffusion Policy (5 Hz) predicts receding-horizon task-space keypoint chunks. A manipulation-centric whole-body controller (50 Hz) tracks them with adaptive end-effector rewards and variable-speed augmentation. Non-vision-grounded keypoints (pelvis/feet) use relative transforms inside the action chunk to limit open-loop drift.

HuMI data collection system and challenges diagram
HuMI hardware, IK preview, and data pipeline. Source: arXiv:2602.06643 HTML figures.

Tasks and success rates

In-domain capability tasks on the paper’s G1 setup (20 trials where stated):

TaskSuccess
Marriage-proposal kneel + ring grasp17/20 (85%)
Unsheathe sword (bimanual)17/20 (85%), 15.7 mm avg EE tracking error
Dynamic toss into container15/20 (75%)
Walk 1–2 m, clean table with lint roller15/20 (75%) with EE+pelvis interface

Ablations bite hard: drop human-in-the-loop kinematic preview on the proposal task 85% → 10%; EE-only interface without whole-body keypoints fails the proposal task 0/10; absolute pelvis tracking on walk-and-clean 75% → 0/10.

Generalization and throughput

For squat-and-pick bottles, authors collect 350 demos across 7 environments and 7 bottle instances. In unseen environments and objects they report 14/20 (70%) success. Versus TWIST2 teleop on the sword task in a 15-minute session: HuMI 62 episodes (96.7% acceptance) vs 28 (64.3% acceptance) — abstract claims data-collection efficiency overall. Proposal kneeling: HuMI 50 demos in 15 minutes; TWIST2 0 usable deep kneels in the same window.

HuMI hierarchical control framework diagram
High-level diffusion policy plus low-level whole-body controller. Source: arXiv:2602.06643.
HuMI data collection throughput comparison with TWIST2
Throughput vs TWIST2 teleop. Source: arXiv:2602.06643 HTML figures.

A Human’s Take

Robot-free whole-body capture only works if the retarget stays honest about object geometry — HuMI’s no-scale + IK preview choice is the right engineering instinct. The TWIST2 comparison on deep kneeling is the money chart: teleop still cannot teach some postures you need for real work.

Sources