Research

BifrostUMI Bridges Robot-Free Demos to G1 Whole-Body Skills

Shar Hendrix 5 min read

Another answer to “how do we collect humanoid whole-body demos without parking a robot in every room”: BifrostUMI from the Beijing Academy of Artificial Intelligence (BAAI). It records portable VR–UMI demonstrations as sparse keypoints, retargets them onto a Unitree G1, and runs a hierarchical controller in the real world.

Paper: arXiv:2605.03452 (v2 July 7, 2026). Project: baai-aether.github.io/BifrostUMI.

Hardware without the humanoid present

Collection uses a Pico 4 motion-capture setup (headset, controllers, two foot trackers, one waist tracker) plus two instrumented handheld grippers with fisheye cameras and magnetic-encoder aperture. Streams include wrist-view RGB, five whole-body keypoints (pelvis, two gripper TCPs, two feet; knees optional), and gripper width. Online Spatial Keypoint Retargeting (SKR) maps human keypoints into robot-native references during capture so operators can see whether the motion stays kinematically sane.

BifrostUMI pipeline from VR-UMI capture to G1 execution
Full BifrostUMI pipeline overview. Source: BifrostUMI project page.

Three-stage control

  1. High-level diffusion policy — DINOv2-encoded wrist views + 15-D lower-body proprioception → 47-D action (five keypoints × 9 + two gripper widths), horizon H = 48.
  2. SKR — preserves metric spatial structure; only scales vertical pelvis-to-foot distance (λ_leg = 0.75 in the paper) for G1 height mismatch; solves constrained whole-body IK.
  3. Low-level whole-body controller50 Hz residual joint tracking trained in MJLab.

Five real G1 tasks

Project demos cover cluttered tabletop pick-and-place, bimanual vegetable collection, dynamic ball shooting, under-table waste disposal, and walking coffee delivery. Ablations on the first three tasks (20 trials each in the paper figures): replacing SKR with GMR hurts pick-place and bimanual success; dropping latency matching hurts timed throws. Seven-keypoint mode (adding knees) matters for deep bends and loco-manipulation.

Throughput vs TWIST2 teleop in 10-minute windows (paper Table I):

OperatorTaskBifrostUMITWIST2Speedup
NoviceBimanual55252.2×
NoviceThrow trash43172.5×
NoviceWalk + coffee61161×
ExperiencedWalk + coffee62512.4×

Average ~2.2× on non-locomotion tasks; loco-manipulation is where robot-free collection opens the widest gap.

BifrostUMI Pico VR trackers and instrumented grippers
Robot-free acquisition hardware. Source: BifrostUMI project page.
BifrostUMI G1 real-world pick-place, bimanual, and ball-throw trials
Real-world G1 evaluations and ablations. Source: BifrostUMI project page.

A Human’s Take

HuMI and BifrostUMI are racing the same bottleneck from different labs: whole-body data without babysitting the robot. The novice coffee-walk number is the one I keep. If your teleop stack cannot teach walking handovers, your training set will never look like a shift.

Sources