Motius reproduction of the TMR architecture trained from scratch for native Unitree G1 motion. The motion input is the canonicalized G1-38D representation at 30 fps: root XY velocity and height, root rotation 6D, and 29 joint angles.

Provenance

Architecture paper: TMR: Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis
Original architecture repository: Mathux/TMR
Motius implementation: TMRBundle / TMRG1Evaluator

Artifact layout

model.safetensors: motion encoder, text encoder, and reconstruction decoder
stats/: G1-38D training-set normalization statistics
config.json and preprocessor_config.json: architecture and protocol
artifact_manifest.json: file and source-checkpoint hashes

The public checkpoint is epoch 139, trained on the 344,966-motion training split of the 359,153-clip HYMotion Data G1 materialization.

Download

python

from motius.evaluation import TMRG1Evaluator

evaluator = TMRG1Evaluator.from_pretrained(
    "ZeyuLing/motius-evaluator-g1-38d-tmr"
)