Hanyang Cao^1,2,*, Yuetong Fang^1,2,*, Taesoo Kwon^3,*, Runyi Yu^2,4, Ji Ma⁵, Jing Tan^1,2,
Yangchen Zhou¹, Baoze Du², Yi Gu¹, Yukang Gao^1,2, Ruoli Dai², Lei Han^2,†, Renjing Xu^1,†

¹HKUST (Guangzhou)   ²Noitom Robotics   ³Hanyang University   ⁴HKUST   ⁵HKU

^*Equal contribution    ^†Corresponding authors

UMR treats the moving exterior body surface as a shared interface between human motion and humanoid robots. It has two main stages:

• Point Cloud Correspondence Learning learns ordered source-robot surface correspondence in aligned canonical poses.
• Correspondence-Guided Retargeting optimizes robot motion using matched surface positions, orientations, contacts, and kinematic constraints.

A learned correspondence is reused by motions with the same source template and target robot.

• September 2026: Added support for the HiPHI dataset. Visit hiphi2smplx to convert HiPHI BVH motions to SMPL-X.

UMR samples the moving exterior surface, so any source with surface-level motion information can be integrated through the same formulation.

For LAFAN1, use lafan_to_smplx to convert BVH motion to SMPL-X before retargeting. Each adapter guide documents the expected local layout.

conda create -n umr python=3.12 pip -y
conda activate umr
python -m pip install --index-url https://download.pytorch.org/whl/cu121 torch==2.4.1
python -m pip install -r requirements-umr.txt

SMPL-X body-model files are not distributed with this repository. Download them from the official SMPL-X provider after accepting its terms. Both .pkl and .npz models are supported; place at least the neutral model at:

smpl/SMPLX_NEUTRAL.pkl
# or
smpl/SMPLX_NEUTRAL.npz

Add SMPLX_MALE and SMPLX_FEMALE in either format when a sequence requires those genders. The NPZ path has been tested with both neutral SMPL-X motion and female OMOMO motion.

The GRAIL example applies its bundled G1-SMPL-X template and pose-corrective overlay to the user-provided neutral SMPL-X model at runtime; the derived baked SMPL-X weights are not distributed.

Retarget the included LAFAN1-derived SMPL-X motion:

python scripts/humanoid_retarget_pipeline.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json

The default configuration uses the included LAFAN1-derived SMPL-X sequence sample_data/lafan1_smplx/dance1_subject2.npz. It builds or reuses the learned point-cloud correspondence, runs correspondence-guided retargeting, and opens the MuJoCo viewer.

To use another robot, copy the example config in robot_configs/ and update its name and MJCF path. Prepare the robot T-pose in UMR Studio: load the robot asset folder, select its MJCF, adjust it into a T-pose, and click Copy T-pose Config. Paste the copied tpose_qpos into the new robot config, then run the pipeline with that config. No manual human-robot mapping is required.

The same surface-based formulation is exposed for other motion representations and interaction settings:

# BONES-SEED SOMA motion
python scripts/humanoid_retarget_pipeline.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json \
  --defaults humanoid_retarget_defaults_bones_seed.json

# Humanoid Character spin-kick
python scripts/humanoid_retarget_pipeline_character.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json

# GRAIL human-scene interaction
python scripts/humanoid_retarget_pipeline_hsi_hoi.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json \
  --defaults humanoid_retarget_defaults_hsi_hoi_grail.json

# OmniContact human-object interaction (pre-converted SMPL-X input)
python scripts/humanoid_retarget_pipeline_hsi_hoi.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json \
  --defaults humanoid_retarget_defaults_hsi_hoi_standard.json

# OMOMO human-object interaction
python scripts/humanoid_retarget_pipeline_hsi_hoi.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json \
  --defaults humanoid_retarget_defaults_hsi_hoi_standard.json \
  --data sample_data/omomo \
  --seq-key sub1_plasticbox_015

# NR FBX/BVH human motion
python scripts/humanoid_retarget_pipeline_nr.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json

# AdaPT body+racket correspondence and retargeting
python scripts/humanoid_retarget_pipeline_adapt.py

# HiPHI SMPL-X motion
python scripts/humanoid_retarget_pipeline_hiphi.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json \
  --data sample_data/hiphi/data/Getting_up/get_up/Getting_up-get_up_0004

See the HiPHI adapter guide for the converted dataset layout, shared correspondence behavior, interaction objects, and batch usage.

Open a result with the default GLFW viewer:

python scripts/visualize_robot_retarget_result.py \
  --result output/unitree_g1_retarget/dance1_subject2_smplx_unitree_g1.npz \
  --play

For browser or headless batch visualization, point the Viser backend at a result folder:

python scripts/visualize_robot_retarget_result.py \
  --result-dir output/unitree_g1_retarget \
  --viewer-backend viser \
  --viser-host 0.0.0.0 \
  --viser-port 8080 \
  --play

Open the clickable Network URL printed in the terminal.

Use Refresh to load new results while batch retargeting is running.

Run the SMPL-X/LAFAN batch pipeline with:

python scripts/humanoid_retarget_pipeline_batch.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json \
  --batch-config humanoid_retarget_defaults_batch.json

BONES-SEED uses its own batch defaults:

python scripts/humanoid_retarget_pipeline_batch.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json \
  --batch-config humanoid_retarget_defaults_batch_bones_seed.json

For a complete converted HiPHI dataset, the dedicated batch runner prepares missing convex object assets and retargets all discovered sequences:

python scripts/humanoid_retarget_pipeline_hiphi_batch.py \
  --config robot_configs/humanoid_retarget_unitree_g1_example.json

Its object preprocessing and retargeting stages default to one worker to bound peak memory. Increase --object-workers or --workers explicitly when the machine has enough memory. HiPHI batch results are saved under output/batch_retarget_hiphi/<robot-name>/.

Add --motion-folder sample_data/bones-seed/motions_proportional/bvh for the actor-proportional subset. BONES-SEED associates each Axxx motion with its matching shape and reuses one correspondence per source-template/robot pair.

Batch defaults use bidirectional warm start with dynamic programming to reduce sensitivity to occasional singularities. This mode is recommended for large-scale retargeting.

Results are saved under output/batch_retarget/<robot-name>/; batch_summary.json records each clip status.

--config selects the target robot. Robot-specific tpose_qpos, joint limits, and model paths belong in this file. --defaults selects source- and task-specific settings.

Surface weights are defined on the motion source, not per robot:

Each segment specifies sample_slots, point_cost, and normal_cost. Robots sharing the same source/task use the same values; only the robot config changes. Interaction defaults give more weight to end-effector preservation. These settings work well for G1 and generally transfer to other robots, but may not be optimal for every embodiment.

• For HSI/HOI, convex-decompose concave objects with CoACD before retargeting. MuJoCo treats a single mesh collision geom as its convex hull.
• BONES-SEED motion and SOMA-X asset placement is documented in the BONES-SEED guide; py-soma-x is included in the requirements.
• NR FBX/BVH input requires Node.js 18 or newer. The bundled minimal Three.js code is used only for FBX mesh parsing, not visualization.

If you find UMR useful in your research, please cite the paper:

@misc{cao2026unifiedmotionretargetinghumanoids,
  title={Unified Motion Retargeting for Humanoids with Learned Point Cloud Correspondence},
  author={Hanyang Cao and Yuetong Fang and Taesoo Kwon and Runyi Yu and Ji Ma and Jing Tan and Yangchen Zhou and Baoze Du and Yi Gu and Yukang Gao and Ruoli Dai and Lei Han and Renjing Xu},
  year={2026},
  eprint={2609.02134},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2609.02134},
}