Project Overview
As a project for a first-semester HfK course, the vision was to explore emergent AI behavior by training physics-based agents to master complex physical tasks. The goal was not just to have an AI solve a parkour course, but to create agents that could develop their own unique and robust locomotion strategies to navigate unpredictable environments and even participate in simple games like soccer or tag.
Gallery
Tech Stack
- Unity
- C#
- Unity ML-Agents
- Python
- TensorFlow
Key Features
- Complex Parkour Environment: A multi-stage parkour course was built in Unity, featuring dynamic obstacles like cannons and moving platforms.
- Robust Agent Training: Using reinforcement learning, a humanoid agent was trained to navigate the course from start to finish, learning resilient and adaptive movements.
- Diverse Skill Sets: Beyond the main parkour course, agents were successfully trained on several other objectives, including a basic version of soccer and a simple game of catch.
- Emergent Locomotion: A self-taught agent developed a unique, non-human "jump-step" gait that proved to be more stable and effective for withstanding obstacles, showcasing true emergent problem-solving.