---
license: other
license_name: multi-license-see-card
license_link: LICENSE
pretty_name: "Kine2Go: Kinematic Dataset for Unitree Go2 with Diverse Gaits and Motions"
size_categories:
  - n<1K
task_categories:
  - robotics
  - reinforcement-learning
tags:
  - kine2go
  - quadruped
  - unitree-go2
  - motion-retargeting
  - imitation-learning
  - kinematic-dataset
  - mocap
  - genesis-simulator
  - ppo
  - gait
---

# Kine2Go: Kinematic Dataset for Unitree Go2 with Diverse Gaits and Motions

A kinematic motion dataset for the Unitree Go2 quadruped robot. Forty reference clips (dog, horse, and synthetic robot motions) are retargeted to the Go2 morphology and paired with a per-clip imitation-learning policy (PPO) and 20 perturbed rollouts with rendered video. Designed to support training and regularization of behavioral foundation models for legged locomotion (Meta-Motivo style), with secondary applicability to single-clip behavioral cloning and motion-retargeting evaluation.

Paper and pipeline-code links are omitted during anonymous review.

## At a glance

| | |
|---|---|
| Robot | Unitree Go2 (12 actuated joints) |
| Simulator | Genesis |
| Clips | 40 |
| Rollouts per clip | 20 |
| Total rollouts | 800 |
| Reference state dim | 61 (float32) |
| Control / sim FPS | 60 Hz |

## Quick start

```python
import sys; sys.path.insert(0, "scripts")
from load_clip import Clip, list_clips

for name in list_clips("."):
    print(name)

clip   = Clip(".", "ai4_dog_canter")
motion = clip.motion          # np.ndarray (T, 61) float32
config = clip.config          # env / obs / reward / command / policy configs
meta   = clip.metadata        # source, license, fps, n_frames, ...
frames = clip.rollout(0)      # list of per-step dicts (requires torch)
policy = clip.policy_path     # path to the trained PPO checkpoint
```

Validate a downloaded copy:

```bash
python scripts/verify_schema.py
```

## Per-clip layout

```
<clip_name>/
├── motion.npy           # (T, 61) float32 reference trajectory
├── config.json          # env / obs / reward / command / policy configs
├── cfgs.pkl             # same configs, pickle (for tooling)
├── clip.json            # per-clip metadata (source, license, n_frames, ...)
├── traj_NNNN/           # 20 perturbed PPO rollouts (NNNN = 0000..0019)
│   ├── init_state.pkl
│   ├── traj.pkl         # list of per-step frame dicts
│   └── traj.mp4         # rendered video
└── logs/
    └── model.pt         # final PPO checkpoint
```

## Subsets and licensing

Multi-license dataset. Each `clip.json` carries the authoritative per-clip license. Full license texts are in `LICENSES/`.

| Subset | Clip prefix | Clips | Rollouts | License |
|---|---|---|---|---|
| AI4Animation - natural dog mocap | `ai4_dog_*` (excl. `synth`) | 15 | 300 | CC BY-NC 4.0 |
| AI4Animation - synthetic via SIGGRAPH 2018 controller | `ai4_dog_synth_*` | 7 | 140 | CC BY-NC 4.0 |
| Vienna Horse Data Collection (Horse 1, two sessions) | `vhdc_horse1_*` | 12 | 240 | CC BY-SA 4.0 |
| cassi - Solo8 robot motions | `solo8_*` | 6 | 120 | BSD-3-Clause |

## Schema

### `motion.npy` (reference trajectory)

Shape `(T, 61)` float32. Per-frame layout:

| Slice | Field | Description |
|---|---|---|
| `[0:18]`  | `dofs_position` | 6 floating-base DOFs + 12 joint DOFs |
| `[18:36]` | `dofs_velocity` | matching DOF order |
| `[36:48]` | `feet_pos` | 4 feet × xyz, world frame |
| `[48:51]` | `base_pos` | world-frame xyz |
| `[51:55]` | `base_quat` | scalar-first quaternion (w, x, y, z) |
| `[55:58]` | `base_lin_vel` | linear velocity |
| `[58:61]` | `base_ang_vel` | angular velocity |

### `traj_NNNN/traj.pkl` (rollouts)

A `list` (one entry per simulator step). Each entry is a `dict[str, torch.Tensor]` with the following keys:

| Key | Shape | dtype | Description |
|---|---|---|---|
| `dof_pos` | (1, 12) | float32 | joint angles |
| `dof_vel` | (1, 12) | float32 | joint velocities |
| `base_quat` | (1, 4) | float32 | base orientation, scalar-first quaternion |
| `base_ang_vel` | (1, 3) | float32 | base angular velocity |
| `actions` | (1, 12) | float32 | PPO policy output |
| `frame` | (1,) | int32 | episode-relative simulator step counter |
| `links_pos` | (17, 3) | float32 | world-frame xyz per link |
| `links_rot` | (17, 6) | float32 | 6-d continuous rotation per link |

### Conventions

- Z-up coordinates (X forward, Y left, Z up).
- Linear units: meters; angular units: radians; quaternions are scalar-first.
- Joint order: `FR_hip, FR_thigh, FR_calf, FL_hip, FL_thigh, FL_calf, RR_hip, RR_thigh, RR_calf, RL_hip, RL_thigh, RL_calf`.
- Feet order (in `motion.npy[36:48]`): `FL_foot, RL_foot, FR_foot, RR_foot`.

The full machine-readable schema is in `metadata.json`.

## Collection and preprocessing

Source motions were retargeted to the Go2 12-DOF morphology and used to train per-clip imitation policies in the Genesis simulator:

- **AI4Animation - natural** (Zhang et al., SIGGRAPH 2018): dog mocap data.
- **AI4Animation - synthetic**: a custom trajectory collector was injected into the AI4Animation SIGGRAPH 2018 controller to drive the trained model along authored paths (circle, ellipse, figure-eight, square, strafes, and a half-flipping jump).
- **Vienna Horse Data Collection**: biomechanical horse kinematics across two measurement sessions × {walk, trot} × 3 repetitions.
- **cassi - Solo8** (Li et al., ICRA 2023): six trajectories from cassi's Solo8 motion bank (slow/fast crawl, scoot, walk, two jump variants).

Each retargeted reference was used to train a PPO policy (`rsl_rl_lib`) in Genesis with hyperparameters captured in `config.json`. After training, 20 perturbed rollouts were recorded per clip.

## Citations

If you use this dataset, please cite the Kine2Go paper *and* the upstream sources for whichever subsets you use.

```bibtex
@inproceedings{kine2go2026,
  title     = {Kine2Go: Kinematic Dataset for Unitree Go2 with Diverse Gaits and Motions},
  author    = {<authors TBD>},
  booktitle = {NeurIPS Datasets and Benchmarks Track},
  year      = {2026}
}

% AI4Animation - applies to ai4_dog_* and ai4_dog_synth_*
@article{zhang2018modeadaptive,
  title   = {Mode-Adaptive Neural Networks for Quadruped Motion Control},
  author  = {Zhang, He and Starke, Sebastian and Komura, Taku and Saito, Jun},
  journal = {ACM Transactions on Graphics},
  year    = {2018}
}

% Vienna Horse Data Collection - applies to vhdc_horse1_*
@misc{vhdc,
  title = {Vienna Horse Data Collection (VHDC)},
  url   = {https://horse.cs.uni-bonn.de/vhdc-home.html}
}

% cassi - applies to solo8_*
@inproceedings{li2023versatile,
  title     = {Versatile Skill Control via Self-supervised Adversarial Imitation of Unlabeled Mixed Motions},
  author    = {Li, Chenhao and Blaes, Sebastian and Kolev, Pavel and Vlastelica, Marin and Frey, Jonas and Martius, Georg},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2023}
}
```