kine2go-pipeline

Quadruped motion retargeting and imitation learning for the Unitree Go2.

Installation · Quick Start · Pipeline · Stage details · Tools

---

kine2go-pipeline retargets animal and quadruped motion capture (AI4Animation dog, Vienna Horse Data Collection, Solo8) into Unitree Go2-executable reference trajectories, and trains PPO policies in Genesis to imitate them. The artifacts produced with this pipeline are published as the kine2go dataset on HuggingFace.

At a Glance

Robot	Unitree Go2 (12 DoF, 4 feet)
Simulator	Genesis 0.3.10
RL	PPO via rsl-rl-lib 3.0.0
Source datasets	AI4Animation, VHDC, Solo8
Python	≥ 3.12
Dataset	kine2go on HuggingFace

Pipeline

  raw mocap (.txt / .csv / .pt)
            │
            ▼  motion_retargeting/main.py
  retargeted trajectory (.npy)
            │
            ▼  motion_imitation/imitation.py
  trained PPO policy (model_*.pt)
            │
            ▼  tools/gather_trajectories.py
  rollout dataset

• motion_retargeting/ — converts source mocap into a Go2 reference trajectory by solving per-frame inverse kinematics in Genesis.
• motion_imitation/ + go2_genesis/ — RL environment, training, and evaluation. motion_imitation adds the imitation reward and observation wrapper on top of the Go2 locomotion env in go2_genesis.
• tools/ — scripts for collecting and post-processing rollout data with a trained policy.

Dataset

Pipeline artifacts — retargeted clips, trained policies, rollouts, and videos — are published at https://huggingface.co/datasets/MIMUW-Robotics/kine2go. Clip names in motion_imitation/motion_ranges.txt mirror the dataset's per-clip folder names.

Installation

git clone <repo-url> && cd kine2go-pipeline
uv sync

uv picks up the pinned Python version from .python-version and resolves all dependencies declared in pyproject.toml.

Note: Genesis runs RL training on a CUDA-capable GPU. Retargeting uses the CPU backend and works without one. Pass --cpu to motion_imitation/imitation.py for a CPU training smoke-test.

Quick Start

End-to-end example using the canonical ai4_dog_walk_00 clip from motion_imitation/motion_ranges.txt (frames 101–580 of dog_walk00_joint_pos.txt):

# 1. Retarget AI4Animation dog walk to Go2
uv run -m motion_retargeting.main \
  --motion-path motion_retargeting/data/AI4Animation/dog_walk00_joint_pos.txt \
  --dataset-name ai4animation \
  --frame-start 101 --frame-end 580
# → motion_retargeting/results/dog_walk00_joint_pos.npy

# 2. Train an imitation policy on the retargeted clip
uv run -m motion_imitation.imitation ai4_dog_walk_00 \
  motion_retargeting/results/dog_walk00_joint_pos.npy \
  --num-envs 4096 --max-iterations 1000
# → logs/ai4_dog_walk_00/model_*.pt

# 3. Evaluate and record a video
uv run -m motion_imitation.imitation_eval logs/ai4_dog_walk_00/model_1000.pt --record
# → logs/ai4_dog_walk_00/recording.mp4

Add --wandb-mode offline (or disabled) for runs without a Weights & Biases account.

Stage details

Retargeting (motion_retargeting/)

uv run -m motion_retargeting.main \
  --motion-path <path>  --dataset-name {ai4animation,horse,solo8}  \
  [--frame-start N] [--frame-end N]  \
  [--scene.record-video] [--robot.<field>=...] [--scene.<field>=...]

Source data lives in motion_retargeting/data/<source>/. Output is written to motion_retargeting/results/<motion_name>.npy, where <motion_name> is the source filename stem. Pass --scene.record-video to also dump an MP4 to motion_retargeting/videos/<motion_name>/.

Imitation training (motion_imitation/ + go2_genesis/)

uv run -m motion_imitation.imitation EXP_NAME MOTION_PATH \
  [--motion-start N] [--motion-end N] \
  [--num-envs 4096] [--max-iterations 1000] \
  [--wandb-mode {online,offline,disabled}] [--cpu]

motion_imitation/motion_ranges.txt is the catalogue of named clips with their frame ranges; names follow the kine2go HuggingFace dataset convention (ai4_dog_*, solo8_*, vhdc_horse1_*). Checkpoints land in logs/<exp_name>/, and the reference clip is copied to logs/<exp_name>/motion.npy so evaluation can reconstruct the frame range automatically.

Evaluation

uv run -m motion_imitation.imitation_eval CKPT_PATH \
  [--record] [--headless] [--num-episodes 1]

--record writes logs/<exp>/recording.mp4.

Tools

Scripts in tools/ for working with rollouts:

• gather_trajectories.py — roll out a trained policy and save trajectories.
• cut_trajectory.py — trim a recorded trajectory to a frame range.
• visualize_trajectory.py — render an MP4 of a recorded trajectory.

Project layout

kine2go-pipeline/
├── motion_retargeting/                 # stage 1: source mocap → Go2 trajectory
├── motion_imitation/                   # stage 2: imitation env, training, eval
│   └── motion_ranges.txt               # catalogue of clip names + frame ranges
├── go2_genesis/                        # Go2 locomotion env, RL training infra
├── tools/                              # rollout collection and post-processing
└── pyproject.toml

Citation

If you use our work, please cite:

@misc{pałucki2026kine2gokinematicdatasetunitree,
      title={Kine2Go: Kinematic dataset for the Unitree Go2 robot with diverse gaits and motions}, 
      author={Władysław Pałucki and Paweł Siwak and Krzysztof Ciebiera and Marek Cygan},
      year={2026},
      eprint={2606.14433},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.14433}, 
}