🐾 TARLoco: Teacher-Aligned Representations via Contrastivde learning for Quadrupedal Locomotion

August 20, 2025 · View on GitHub

Official implementation of our IROS 2025 paper | 🌐 Project Website with videos and results

TARLoco is a modular and scalable framework for blind quadrupedal locomotion that integrates:

Teacher–Student training with representation learning via contrastive loss.
High‑fidelity simulation using Isaac Sim and streamlined workflows via Isaac Lab.
Sim‑to‑Real deployment on Unitree Go2 (SDK2 + LCM).

🔥 News

[18-08-2025]: The project website with paper, videos, and results is live.
[15-06-2025]: This work is accepted at IROS 2025.
[---Soon---]: Sim2Real module (scripts + pretrained checkpoints) will be added here shortly. Stay tuned!

Robust blind locomotion must adapt to unknown dynamics (e.g., mass, friction, terrain). TARLoco aligns a student’s proprioceptive representation with a teacher’s privileged latent using contrastive learning, yielding policies that are:

Generalizable across non‑stationary settings.
Compact & deployable on real hardware.
Interpretable via t‑SNE / UMAP visualizations of embeddings.

🛠 Installation

Tested on: Ubuntu 22.04 • CUDA 12.1 • Python 3.10

# 1) Conda env
conda create -n tar python=3.10 -y
conda activate tar

# 2) Clone the repo
git clone https://github.com/ammousa/TARLoco.git
cd TARLoco

# 3) PyTorch (CUDA 12.1)
pip install --upgrade pip
pip install torch==2.5.1 --index-url https://download.pytorch.org/whl/cu121

# 4) Isaac Sim SDK
pip install isaacsim[all,extscache]==4.5.0 --extra-index-url https://pypi.nvidia.com

# 5) Isaac Lab
git clone --branch v2.1.0 https://github.com/isaac-sim/IsaacLab.git _isaaclab
sudo apt install -y cmake build-essential
./_isaaclab/isaaclab.sh --install

# 6) Smoke test (headless sim)
python _isaaclab/scripts/tutorials/00_sim/create_empty.py --headless

# 7) Dev install
pip install -e .

Tip: If you use W&B logging, run wandb login once before training.

🚀 Usage

Framework Tasks

We offer a comprehensive suite of tasks for training and evaluation, spanning various architectures and baselines:

Category	Training Task	Evaluation Task
TAR (MLP Architectures)	`go1-train-tar-mlp-rough`	`go1-eval-tar-mlp-rough`
	`go1-train-tar-mlp-no-priv-rough`	`go1-eval-tar-mlp-no-priv-rough`
	`go1-train-tar-mlp-no-priv-no-vel-rough`	`go1-eval-tar-mlp-no-priv-no-vel-rough`
TAR (RNN Architectures)	`go1-train-tar-rnn-rough`	`go1-eval-tar-rnn-rough`
	`go1-train-tar-rnn-no-priv-rough`	`go1-eval-tar-rnn-no-priv-rough`
	`go1-train-tar-rnn-no-priv-no-vel-rough`	`go1-eval-tar-rnn-no-priv-no-vel-rough`
TAR (TCN Architectures)	`go1-train-tar-tcn-rough`	`go1-eval-tar-tcn-rough`
SLR	`go1-train-slr-rough`	`go1-eval-slr-rough`
HIMLoco	`go1-train-him-rough`	`go1-eval-him-rough`
Expert Teacher Policies	`go1-train-teacher-rough`	`go1-eval-teacher-rough`
	`go1-train-teacher-encoder-rough`	`go1-eval-teacher-encoder-rough`
	`go1-train-teacher-rnn-rough`	`go1-eval-teacher-rnn-rough`

Below are examples of how to train and evaluate TARLoco models. You can replace the --task argument with any task from the table above to experiment with different configurations.

Training Example

Train a TAR student (RNN) on rough terrain:

python standalone/tarloco/train.py \
  --task go1-train-tar-rnn-rough \
  --headless --device cuda

Evaluation Example

Evaluate a checkpoint for the same task but with additional evaluation conditions:

python standalone/tarloco/evaluate.py \
  --task go1-eval-tar-rnn-rough \
  --headless --device cuda \
  --load_run 2025-07-25_00-54-10_SLR_S96_29 \
  --checkpoint model_10000

Results

We host full results on 📊 W&B Dashboard. Please check our 🌐 project website for more details and testing videos.

🌟 Acknowledgments

Special thanks to Bruno Adorno and Amy Johnson from the University of Manchester for providing the hardware for testing and for their unwavering support throughout the project.

This work builds upon:

IsaacLab - Simulation backbone
RSL-RL - RL foundation
SLR & HIMLoco - Baseline comparisons

These components retain their original licenses -see licenses/-. Only new contributions in this repository are licensed under CC BY-NC-SA 4.0.

🔗 Citation

If you use TARLoco in your research, please cite:

@misc{mousa2025tarteacheralignedrepresentationscontrastive,
      title={TAR: Teacher-Aligned Representations via Contrastive Learning for Quadrupedal Locomotion}, 
      author={Amr Mousa and Neil Karavis and Michele Caprio and Wei Pan and Richard Allmendinger},
      year={2025},
      eprint={2503.20839},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2503.20839}, 
}

⭐ Contributing

We welcome issues and PRs! If this project helps your research, please ⭐ the repo and share feedback.

Unlock the Future of Blind Locomotion
⭐ Star the Repository • 📄 Read the Paper • 🎬 Watch the Demo