On a tennis clip, a Unitree G1 policy trained on a three-camera capture stayed up in all 59 randomised physics runs and tracked within 0.6 cm of one trained on a 27-camera mocap dome (12.9 vs 12.3 cm keypoint error). Trained on single-camera estimates of the same clip it still stayed up but drifted 3–6 cm further. Across 10 clips of a 15-clip study, the three-camera version completed 551 of 590 runs; five more clips are being re-scored.
01Setup
We took clips from the CMU Panoptic dome, which films people with dozens of synchronised cameras, and built the same motion three ways: from the dome’s dense 3D reconstruction, from three of its cameras run through our pipeline, and from one camera with two leading single-view estimators. Each version was retargeted to a Unitree G1 with GMR and tracked by a motion-tracking policy trained in physics with the same recipe and seed. Each policy then ran 60 times in plain MuJoCo with G1 torque limits: one nominal run and 59 with random pushes, floor friction, mass and motor-gain changes and control delays. Panoptic and the single-camera estimators are used here for evaluation only; their licences do not allow commercial use.
02Result on a tennis clip
| Capture used for training | Runs that stay up | Keypoint error | Bone-direction error |
|---|---|---|---|
| 27-camera dome reconstruction | 59 / 59 | 12.3 cm | 9.6° |
| Three cameras (our pipeline) | 59 / 59 | 12.9 cm | 12.5° |
| One camera, estimator A | 59 / 59 | 18.1 cm | 15.3° |
| One camera, estimator B | 59 / 59 | 15.5 cm | 13.0° |
Keypoint error is measured against the dome’s reconstruction of the person. Robustness no longer separates the sources once the retargeting is clean; position and root accuracy do.
03Across clips
Across a 15-clip study, on the 10 clips where the three-camera capture followed the target person, policies trained on it completed 551 of 590 randomized runs; policies trained on the dome’s reconstruction completed 474 of 590. Where the two captures agree (2.7–4.2 cm apart), a policy can still fail on one and succeed on the other, which points to training variance rather than the capture. Five sport and dance clips are being re-scored: their dome reference turned out to be 0.7 s out of sync with the video. Once aligned, the three-camera capture agrees with the dome to 5.8–8.0 cm on them.
04What made robots fall
Early in the study, most falls traced to defects in the data rather than the cameras: a start-up transient in retargeting, joints no camera keypoint observes, and frozen frames where tracking gaps were filled. Flagging those per clip made the three-camera capture as robust as the dome. This matches the State of AI Report 2026’s point that quality labels decide whether extra data helps or hurts.
05Limits
- Simulation only; no real-robot runs yet.
- One random seed per configuration; seed-to-seed spread was up to six runs out of 59 on an earlier version.
- The dome reconstruction is itself markerless, so this compares capture setups, not absolute truth.
Questions
Do you need a mocap studio to train a humanoid from human motion?
Not for the motions we tested: three ordinary cameras gave the same robustness as a 27-camera dome and tracking within 0.6 cm on a tennis clip.
How much worse is a single camera?
On the same clip, single-camera estimates added 3–6 cm of keypoint error to the trained robot’s tracking, though the robot still stayed up.
Was this tested on a real robot?
Not yet. All runs are in MuJoCo physics with randomised disturbances.