DiMorph Body-Field Encoder

Runtime checkpoint for the DiMorph cross-morphology body functional-field encoder used by T1, K1, Unitree H1 and Unitree G1 policies.

The public artifact is a sanitized inference checkpoint. Training-only optimizer/pretrainer state, source file lists, cached human roles, normalization statistics and raw/converted motion data are not included. The runtime keys required by research/dimarl/dimarl/model.py are preserved.

Architecture

  • encoder type: body_functional_field_v1
  • spectral coordinate width: 32
  • hidden width: 128
  • role width: 64
  • attention heads: 4
  • encoder layers: 2
  • source morphology joint count: 23 (Booster T1)

Download

hf download tsingzeyong/dimarl-body-field-encoder \
  encoder_best_runtime.pt \
  --local-dir /root/gpufree-data/DI-MARL/body_field/public

Use it during training with:

export DIMARL_PRETRAINED_ENCODER=/root/gpufree-data/DI-MARL/body_field/public/encoder_best_runtime.pt

The full code and experiment record are available at hashlure/di-marl, commit 3446d70bdf1e2cd051e60288f39e068ad3d973ce.

Validation status

The GitHub snapshot passed 99 CPU static tests, Python AST checks and shell syntax checks. The post-fix four-robot GPU training sequence has not yet been rerun, so this repository does not claim final locomotion performance.

Terms

The encoder was trained for non-commercial scientific research using AMASS/CMU-derived supervision. Use is limited by the applicable source-dataset terms, including the AMASS license. No AMASS/CMU frames are distributed here.

Downloads last month
-
Video Preview
loading