DOGE : Dynamic Obstacle Ground-based Evasion

This repository contains a complete research and experimentation framework built on ROS 2 and MuJoCo for learning dynamic projectile evasion with the Solo12 quadruped robot.

The key design goal of this project is high-speed reactivity:

• Perception: Utilizing an event-based camera to track fast-moving objects with microsecond latency.
• Agility: Moving beyond static path planning by using Model-Based Reinforcement Learning (DreamerV3).
• Real-Time: Integrating high-frequency motor control (100Hz) with the neuromorphic vision system.

Table of Contents

• Prerequisites
• Installation
• Usage
• Results & Limitations

Prerequisites

This project assumes familiarity with ROS 2 and Reinforcement Learning.

Before working with this repository, you should:

• Have ROS 2 Humble installed on Ubuntu 22.04
• Have MuJoCo installed for simulation
• Have a Python environment with PyTorch and standard RL libraries

Installation

Clone the repository using the following command:

git clone --recurse-submodules https://github.com/Telios/doge.git

Make a new conda environment and install the required packages using the following commands:

conda env create -f environment.yml
conda activate doge

Install the external libraries

./install_external.sh

Usage

First you need to calibrate the robot.

bash scripts/calibrate.sh

To run doge with Solo12, use the following command:

bash scripts/run_doge.sh

Follow the instructions in the terminal to start doge.

Results & Limitations

Simulation results from the event-based camera

Simulation results in Mujoco

• Simulation Success: The DreamerV3 agent successfully learned a robust policy in simulation, capable of consistently dodging incoming projectiles by coordinating body movement and orientation.

doge.mp4

• Sim-to-Real Challenges: While the policy performed well in simulation, the direct transfer to the physical Solo12 robot faced a significant sim-to-real gap. The real-world physics and sensor noise profiles proved distinct enough that the policy requires further refinement.
• Future Work: Identifying pathways for bridging this gap include more aggressive domain randomization and sophisticated reward shaping to encourage more stable real-world behaviors.