sam-dude/go1-expert-trajectories

🤖 Unitree Go1 Quadruped Expert Trajectory Dataset

This dataset contains expert trajectory demonstrations for the Unitree Go1 quadruped robot operating in Webots. It serves as the initialisation dataset for Behavior Cloning ($\pi_{BC}$) in the paper:

"Training Walking Robots on Low-Cost Hardware: A CPU-Only Approach for African Urban Deployment"

---

📊 Dataset Structure & Contents

The dataset is packaged in NumPy compressed archive format (.npz):

• trajectory_data.npz (1.3 MB):

  • observations: (10000, 36) array containing normalized robot states (roll, pitch, angular velocities, joint positions, joint velocities).
  • actions: (10000, 12) array containing target joint angle motor commands for the 12 Go1 leg actuators.

• trajectory_data_window3.npz (3.25 MB):

  • observations: (10000, 108) array incorporating a 3-step temporal observation window for sequence modeling.
  • actions: (10000, 12) array containing target joint angle motor commands.

---

💻 Usage & Loading in Python

Using NumPy

import numpy as np

# Load trajectory dataset
data = np.load("trajectory_data.npz")

observations = data["observations"] # Shape: (10000, 36)
actions = data["actions"]           # Shape: (10000, 12)

print(f"Loaded {len(observations)} expert state-action pairs.")

Downloading via Hugging Face Hub

from huggingface_hub import hf_hub_download
import numpy as np

file_path = hf_hub_download(
    repo_id="sam-dude/go1-expert-trajectories",
    filename="trajectory_data.npz",
    repo_type="dataset"
)

data = np.load(file_path)
print("Dataset loaded successfully!")

---

📜 Citation

@inproceedings{ibiyemi2026cpu,
  title={Training Walking Robots on Low-Cost Hardware: A CPU-Only Approach for African Urban Deployment},
  author={Ibiyemi, Samuel and Akinremi, Bunmi},
  year={2026}
}