Humanoid Manipulation via Imitation Learning

🏆 1st Place — ETH Robotics Club HACK2026

Overview

This repository documents my contribution to a humanoid robot system developed during ETHRC HACK2026.

As part of a 6-person team, I focused on the manipulation pipeline, including teleoperation data collection (Homunculus Exoskeleton) and training an imitation learning–based grasping policy using LeRobot.

The full system was built collaboratively, with separate components for locomotion and navigation (linked below). This repository provides an overview of my work, along with results, data, and references to the complete system.

Method

• Imitation learning with LeRobot
• ~1.5h teleoperation data Homunculus Exoskeleton
• ACT policy for object grasping

Results

• Trained on a single object (Coke can)
• Successfully generalized to objects with different sizes and appearances

🎥 Demo video:

📊 Presentation: Slides

Repository Structure

• policy/ – trained model
• data/ – demonstration data
• results/ – videos / outputs

System Components

• Manipulation

  • 🤗 Dataset: g1-dex3-pick
  • 🤗 Model: g1-dex3-pick-act-100k

• Locomotion (GR00T WBC)

  • https://github.com/mglhbr/GR00T-WholeBodyControl

• Navigation & GPS

  • https://github.com/mglhbr/ethrc_hack_26

Acknowledgements

Huge thanks to my teammates and ETH Robotics Club for organizing the hackathon.