UniSkill: Imitating Human Videos via Cross-Embodiment Skill Representations

Hanjung Kim*, Jaehyun Kang*, Hyolim Kang, Meedeum Cho, Seon Joo Kim, Youngwoon Lee

[arXiv][Project][Dataset][BibTeX]

Features

• A universal skill represenation (UniSkill) that enables cross-embodiment imitation by learning from large-scale video datasets, without requiring scene-aligned data between embodiments.
• UniSkill supports skill transfer across agents with different morphologies, incluiding human-to-robot and robot-to-robot adaptation.
• UniSkill leverages large-scale video pretraining to capture shared interaction dynamics, enhancing adaptability to novel settings.

Todos

We will be releasing all the following contents:

• FSD & ISD training code
• FSD & ISD checkpoint
• Real-world Dataset
• Skill-Conditioned Policy training & inference code
• Skill-Conditioned Policy checkpoint
• Skill-extraction code

Installation

• Linux or macOS with Python ≥ 3.10
• PyTorch ≥ 2.3 and torchvision that matches the PyTorch installation. Install them together at pytorch.org to make sure of this.
• pip install -r requirements.txt

Getting Started

For dataset preparation instructions, refer to Preparing Datasets for UniSkill.

We provide the script train_uniskill.py for training UniSkill.

Training UniSkill (FSD & ISD)

To train the Forward Skill Dynamics (FSD) and Inverse Skill Dynamics (ISD) models of UniSkill, first set up your custom datasets. Once your dataset is ready, run the following command:

cd diffusion

python train_uniskill.py \
    --do_classifier_free_guidance \
    --pretrained_model_name_or_path timbrooks/instruct-pix2pix \
    --allow_tf32 \
    --train_batch_size 32 \
    --dataset_name {Your Dataset} \
    --output_dir {output_dir} \
    --num_train_epochs 50 \
    --report_name {report_name} \
    --learning_rate 1e-4 \
    --validation_steps 50

Multi-GPU Training

For multi-GPU training, first modify the configuration file hf.yaml as needed. Then, run the following command:

accelerate launch --config_file hf.yaml diffusion/train_uniskill.py \
    --do_classifier_free_guidance \
    --pretrained_model_name_or_path timbrooks/instruct-pix2pix \
    --allow_tf32 \
    --train_batch_size 32 \
    --dataset_name {Your Dataset} \
    --output_dir {output_dir} \
    --num_train_epochs 50 \
    --report_name {report_name} \
    --learning_rate 1e-4 \
    --validation_steps 50

Make sure to replace {Your Dataset}, {output_dir}, and {report_name} with the appropriate values.

Download

• FSD & ISD

• Skill-conditioned Diffuion Policy

Citing UniSkill

@inproceedings{
    kim2025uniskill,
    title={UniSkill: Imitating Human Videos via Cross-Embodiment Skill Representations},
    author={Hanjung Kim and Jaehyun Kang and Hyolim Kang and Meedeum Cho and Seon Joo Kim and Youngwoon Lee},
    booktitle={9th Annual Conference on Robot Learning},
    year={2025}
}