README.md

---
license: unknown
pretty_name: CMR Retargeted Motions
tags:
- robotics
- humanoid
- motion-retargeting
- reinforcement-learning
- mujoco
size_categories:
- n<1K
---

# CMR Retargeted Motions

Motions retargeted to the Unitree G1 (29 DoF) humanoid and exported to the RL format the holosoma
stack consumes. This dataset is compliant with the holosoma motion-retargeting RL training pipeline.

The files are produced by the [`sqp_retargeting`](https://github.com/jonarriza96/sqp_retargeting)
repo (`export/convert_data_format_mj.py`), which replays already-retargeted `qpos` through MuJoCo
forward kinematics and records per-body world kinematics. Each timestamped run folder holds one
subfolder per suite (`robot_only_omomo/`, `robot_object_omomo/`, `robot_terrain/`), each with one
compressed `.npz` per clip, plus the run's `comparison.md`/`comparison.json`.

## Download

```bash
hf download jonarriza96/cmr_data --repo-type dataset --local-dir ./data
```

This downloads the run folders into `data/`, skipping files already present. The dataset is
**private**, so first `pip install huggingface_hub`, get access on Hugging Face, and authenticate once:

```bash
hf auth login   # token from https://huggingface.co/settings/tokens
```

## Train

Run the RL training with the corresponding holosoma command, pointing `motion_dir` at the absolute
path of the downloaded suite folder:

- `robot-only`:

  ```bash
  python src/holosoma/holosoma/train_agent.py \
      exp:g1-29dof-wbt logger:wandb \
      --command.setup_terms.motion_command.params.motion_config.motion_dir="<abs>/data/102317_170726/robot_only_omomo"
  ```

- `robot-object`:

  ```bash
  python src/holosoma/holosoma/train_agent.py \
      exp:g1-29dof-wbt-w-object logger:wandb \
      --command.setup_terms.motion_command.params.motion_config.motion_dir="<abs>/data/102317_170726/robot_object_omomo"
  ```

## File format

Each `*.npz` (robot-only clip) contains:

| key | shape | dtype | meaning |
|-----|-------|-------|---------|
| `fps` | `(1,)` | int64 | output frame rate |
| `joint_pos` | `(T, 36)` | float64 | generalized position: 3 base pos + 4 base quat (wxyz) + 29 DoF |
| `joint_vel` | `(T, 35)` | float64 | generalized velocity: 3 base lin + 3 base ang + 29 DoF |
| `body_pos_w` | `(T, nbody, 3)` | float64 | per-body world position |
| `body_quat_w` | `(T, nbody, 4)` | float64 | per-body world orientation (wxyz) |
| `body_lin_vel_w` | `(T, nbody, 3)` | float64 | per-body world linear velocity |
| `body_ang_vel_w` | `(T, nbody, 3)` | float64 | per-body world angular velocity |
| `joint_names` | `(29,)` | str | actuated joint names, in `joint_pos`/`joint_vel` order |
| `body_names` | `(nbody,)` | str | MuJoCo body names, in `body_*_w` order |

Object-interaction clips additionally carry `object_pos_w (T,3)`, `object_quat_w (T,4)`,
`object_lin_vel_w (T,3)` and `object_ang_vel_w (T,3)`; for those the object columns are stripped from
`joint_pos`/`joint_vel`.

```python
import numpy as np
d = np.load("sub3_largebox_003_mj_fps50.npz", allow_pickle=True)
joint_pos = d["joint_pos"]   # (T, 36)
```

## License

`unknown` — set this before publishing. These motions are retargeted from upstream sources (e.g.
LAFAN1, OMOMO); the licenses of those datasets govern redistribution.