Jumping Continuously over Discontinuous Terrains

This repository contains the code for the paper Agile Continuous Jumping in Discontinuous Terrains. The name of this repo, Jumping CoD, stands for Jumping Continuously over Discontinuities.

A Jumping Cod

The main contents of the repo includes:

• The simulation environment to train the terrain-aware jumping controller.

• The code to deploy the trained controller to a real Unitree Go1 robot.

• Additional utilities to inspect robot logs and record data for real-to-sim study.

Contents

• Simulation and Training Setup
  • System Requirement
  • Installing Related Packages
  • Running the Demo Policy
• Real Robot Deployment Setup
  • Hardware Preparation
    • 1. Mount the Depth Camera
    • 2. Prepare Remote Estop
    • 3. Mount and Connect the Mac Mini
  • Robot Preparation
  • Mac Mini Software Setup
    • 1. Installing Asahi-Linux
    • 2. Install ROS
    • 3. Clone and Install Repo Dependencies
    • 4. Connect Mac Mini to the real robot
    • 5. Test connection
  • Running the Demo Policy
• Training a New Policy
  • Configuring the Motion Policy
  • Training the Motion Policy
  • Configuring the Heightmap Predictor
  • Training the Heightmap Predictor
• A Few Utility Tools
  • Dog Tracer for Log Inspection
  • Camera Latency Calibration
  • Motor Calibration

Credits

The perception pipeline used in this project is heavily inspired from the Robot Parkour Learning work. More specifically, we design our camera mount based on their 3D design, and used a very similar perception network structure. We thank the authors, especially Zipeng Fu, for his generous support.