AutelicaLearn
Learn.
What we have learned capturing people with ordinary cameras and teaching their motion to a humanoid: the data sources, what camera count changes, what breaks in retargeting, and how the capture works, with live figures on real footage.
Real footage
A 3D skeleton of a sprinter rebuilt from three cameras.
Robot learning
ExplainerHuman motion data for humanoid robots: the six sources and what each one measuresWhere humanoid robots get human motion data: teleoperation, handheld grippers, first-person headcam video, mocap studios, single-camera estimates and markerless multi-view capture compared by what each records, where it works and how accurate it is.ResearchHow much does camera count matter when a humanoid learns from video?We trained a Unitree G1 tracking policy on the same motion captured three ways: a 27-camera mocap dome, three cameras, and one camera. Three cameras matched the dome; one camera added 3–6 cm of keypoint error.EngineeringRetargeting human motion to a Unitree G1: what breaks and how to check itFive failure modes we hit retargeting captured human motion to a Unitree G1 humanoid, from start-up transients to floating feet, and the checks that catch them before training.Research3D hand capture from a few cameras: 5.4 mm on people we never trained onMarkerless 3D hand capture tested leave-one-person-out on the HO-Cap hand-object dataset: median joint error falls from 9.9 mm with off-the-shelf keypoints to 5.4 mm with 8 cameras and 5.6 mm with 4, on people the models never saw.ResearchTraining a humanoid on terrain rebuilt from drone footageWe rebuilt a dirt path from the same three-drone footage that captured a person walking and jogging on it, checked the feet against it, and trained a Unitree G1 tracking policy on that terrain in simulation: 196 of 200 randomised runs stayed up, against 128 for the flat-trained policy.ResearchTraining head-camera hand models with 3D labels from outside camerasThe same head-camera hand-pose model trained on four label sources and scored on people it never saw: labels from four outside cameras came within about 1 mm of ground truth (24.3 vs 23.2 mm), while MediaPipe labels from the head camera alone stalled at 64 mm however much data we added.BenchmarkMeasured vs inferred ground on real slopesOn four real outdoor slopes, ground inferred from a person’s motion alone was about 70 cm off USGS survey lidar at the 90th percentile of foot contacts; ground measured from the same three-drone footage was 10 cm off (3 cm median). On flat grass all methods were within 10 cm.ResearchCan a simulated humanoid climb a rope net?We captured a person climbing a playground rope net with three drones and moved the climb onto a Unitree G1 in simulation. Kinematically the hands and feet stay on the ropes; in physics, on a simplified rigid net, the G1 finishes the climb in 6 of 6 runs with a support force of up to its body weight, and in none with half that.ResearchProcessing a whole afternoon of captureEighteen minutes of one person filmed by three drones, processed end to end rather than as highlight clips: the 3D body is measured on 87–91% of frames, about 95% of the frames two drones can see, and every gap is flagged rather than filled in silently.EngineeringErasing bystanders instead of blurring themField capture in public places films strangers. We stopped blurring them and now remove them from published footage, then check every frame. On 15 public clips, 82% of the footage passed and 18% was cut; our first version missed faint, distant people that a stricter per-frame check found.
How the capture works
Explainer · 4 min · 2 live figuresHow to sync cameras without timecodePhones and action cameras share no clock. How multi-camera footage is lined up from sound and then from the subject’s own motion, how precise each step is, and when it fails. Two live figures.Explainer · 4 min · 3 live figuresOne camera vs. threeWhy a single camera can only estimate 3D motion, what a second camera measures, and what a third adds: fewer blind spots and a built-in accuracy check. Shown live on a real sprint.EngineeringWhy handheld video broke our camera solve, and the one setting that fixed itOur 3D replays from a single handheld phone took 6 to 10 minutes because the camera-path solve failed about 3 runs in 10, nondeterministically. What we ruled out, why rotation-dominated video confuses a global solver, and the fix: 3 minutes, first try.