DEPLOYDatabase

Brain

CLONE

CLONE is a closed-loop whole-body humanoid teleoperation system from Beijing Institute of Technology (BIT), Beijing Institute for General Artificial Intelligence (BIGAI), and Peking University (PKU). It uses a Mixture-of-Experts (MoE) student policy with LiDAR-odometry error correction so an operator can drive coordinated upper- and lower-body motion from only head and wrist tracking on a commercial MR headset (Apple Vision Pro).

The project page and arXiv 2506.08931 report real-world Unitree G1 demos including ground pick-and-place, table tennis strokes, and long-horizon navigation with about 5.1 cm mean straight-path tracking error over 8.9 m. Early deployment checkpoints are released; training and evaluation code remain pending per the project README. Recorded as an unreviewed research-model stub pending deeper editorial pass.

Research model · Maturity: Research · Closed

Machine-readable surfaces

Architecture

MoE student policy (3 layers x 4 experts, top-2 routing) distilled from a privileged teacher; closed-loop global pose via LiDAR odometry (FAST-LIO2) plus Apple Vision Pro operator tracking; trained on augmented AMASS subset (CLONED) with hand-orientation augmentation. Builds on OmniH2O-style teacher-student teleop.

Key facts

What it is
Closed-loop whole-body humanoid teleoperation policy (MoE student + LiDAR odometry error correction). Primaries: project page; arXiv 2506.08931. Institutions: BIT + BIGAI + PKU (authors equal-contribute Yixuan Li, Yutang Lin; corresponding Wei Liang, Yixin Zhu, Siyuan Huang)
Input / hardware
Operator input is head position plus dual-wrist 6D pose from a commercial MR headset (Apple Vision Pro). Real-world demos on Unitree G1; authors acknowledge Unitree Robotics hardware support. Not a commercial product; research teleoperation system
Reported tracking (paper)
Paper reports ~5.1 cm mean straight-path global tracking error (max 12 cm) over 8.9 m trajectories, and long-horizon mixed navigation spanning >15 m with return to start and minimal drift. Cap: author-reported experimental results on project page / arXiv, not independently reproduced
Openness
Early real-world deployment checkpoints and deploy docs released (project README, Jul 2025 news). Training code, evaluation code, and enhanced checkpoints still pending. License noted as CC BY-NC 4.0 on the project GitHub. isOpen=false until full training release is confirmed
Thin-record note
Unreviewed research stub. Academic multi-lab system (BIT/BIGAI/PKU); no registry Company developers attached (labs are not registry Companies). CoRL 2025 venue noted on project materials; deepen architecture / dataset / hardware wiring on a later reviewed pass
CLONED motion dataset
The project page says CLONE trains on a curated CLONED set built by editing a subset of AMASS and adding sampled hand orientations plus further motion capture (CLONE).
Closed-loop sensing stack
The project page says real-world deployment uses LiDAR odometry for humanoid state and Apple Vision Pro tracking so the policy can correct global-pose drift (CLONE).

Common questions

What is CLONE?
CLONE is a closed-loop whole-body humanoid teleoperation system from Beijing Institute of Technology (BIT), Beijing Institute for General Artificial Intelligence (BIGAI), and Peking University (PKU). It uses a Mixture-of-Experts (MoE) student policy with LiDAR-odometry error correction so an operator can drive coordinated upper- and lower-body motion from only head and wrist tracking on a commercial MR headset (Apple Vision Pro). The project page and arXiv 2506.08931 report real-world Unitree G1 demos including ground pick-and-place, table tennis strokes, and long-horizon navigation with about 5.1 cm mean straight-path tracking error over 8.9 m. Early deployment checkpoints are released; training and evaluation code remain pending per the project README. Recorded as an unreviewed research-model stub pending deeper editorial pass.
Is CLONE open source?
No. CLONE is recorded as proprietary on the DEPLOY registry. Model weights and source are not publicly available.
What type of AI is CLONE?
CLONE is a research model, built on a MoE student policy (3 layers x 4 experts, top-2 routing) distilled from a privileged teacher; closed-loop global pose via LiDAR odometry (FAST-LIO2) plus Apple Vision Pro operator tracking; trained on augmented AMASS subset (CLONED) with hand-orientation augmentation. Builds on OmniH2O-style teacher-student teleop. architecture on the DEPLOY registry.
What is CLONE's maturity stage?
CLONE is at the research stage on the DEPLOY maturity ladder. Research stage means active development without commercial deployments on file.
Which robots run on CLONE?
No robot models on the DEPLOY registry are recorded as running CLONE. DEPLOY wires brain-to-model connections only when the wiring is verifiable from primary sources; absence may reflect pre-deployment or unverified manufacturer claims.

Sources (2)

  1. https://humanoid-clone.github.io/ · 2025-06
  2. https://arxiv.org/abs/2506.08931 · 2025-06-10
Methodology: Verified · 2 sources (no primary) · last reviewed 2026-10-09

Verification posture

Verified

Low confidence

Review state

Stable

Last reviewed 2026-10-09

Maturity + lifecycle

Maturity stage: research

Sources by quality tier

1
unclassified
Unclassified source
1
preprint
Preprint

The framework is documented at /methodology. Corrections at /corrections. Reviewer: DEPLOY editorial team.

Methodology surface for CLONE.

Canonical ID 79ebdcaa-5443-43e7-a33b-3e447eba5392