Full-body under-controlled robot retargeting, yo.
FURRY is a student research project on Meta Quest 3 teleoperation and trajectory collection for humanoids in MuJoCo. We start from TWIST2, target Unitree G1 with Dex3-1 hands, and investigate how far sparse human observations can support whole-body motion and dexterous loco-manipulation.
Simulation only. Controller-based and controller-free Quest input are equal research tracks. An external RealSense RGB-D camera is optional. Physical-robot deployment is outside this project's scope.
As of 16 October 2026, the standalone Quest 3 / macOS G1 prototype runs native OpenXR, MuJoCo and ONNX control on the headset. Meta body tracking passes through C++ GMR; controller wrist targets, passthrough, calibration, policy selection, trajectory capture and operator MP4 recording are available. TWIST2, HOMIE and SONIC modes and experimental fine-tuned policies are included.
The final demo package shows tracking, cup,
Kick-T and tennis, with a Rubik teaser. Published Quest datasets and autonomous
visual-policy results are available in datasets/ and the prototype's training
reports. These demonstrations do not establish general task reliability:
walking, external-camera leg tracking and arbitrary Rubik sequences remain
research limitations. Physical-robot deployment remains outside scope.
docs/ is authoritative; the Wiki is a manually published mirror. See the
publishing procedure.
Upstream software, models and datasets have their own licenses. FURRY's license does not relicense them. The Quest training snapshot includes four TWIST2 example-motion replay fixtures with their original MIT notice; see its dataset README for provenance.