Colosseum
Colosseum is a research playground for humanoid robot soccer, built on top of mjlab — a GPU-accelerated reinforcement learning framework for robotics powered by MuJoCo Warp.
The primary platform is the Booster T1 humanoid (23 DOF full body). Training runs thousands of parallel environments on a single GPU, enabling fast iteration on complex locomotion and ball-manipulation tasks.
Key features
- Massive parallelization — 4096+ environments simultaneously on GPU via MuJoCo Warp
- Manager-based configuration — rewards, observations, actions, and events declared as composable term configs; no boilerplate loops
- Shared training / deployment code — pure MDP functions are used unchanged in both training and on-robot inference
- Curriculum support — curricula built into the task layer
- Research-ready — each project lives in its own
research/folder with a dedicated training pipeline and evaluation protocol
Where to start
| I want to… | Go to |
|---|---|
| Install and run a first training | Setup → Getting Started |
| Understand the framework internals | mjlab Overview |
| Build a new task from scratch | CartPole tutorial |
| Read about dribbling with obstacle avoidance | Dribbling |
| Read about goal-conditioned maze navigation | Soccer Maze |