Robotic dinosaur locomotion research using reinforcement learning and MuJoCo physics simulation.
Mesozoic Labs is a research project exploring bipedal and quadrupedal locomotion in robotic dinosaurs. We use MuJoCo for realistic physics simulation and train agents with algorithms like PPO and SAC.
Goals:
- Develop realistic locomotion controllers for various dinosaur species
- Explore predatory behaviors (hunting, striking, pack coordination)
- Create transferable policies for robotic applications
- Experiment with JAX/MJX for high-performance training
mesozoic-labs/
├── environments/ # Dinosaur training environments
│ ├── velociraptor/ # Velociraptor (bipedal predator with sickle claws)
│ │ ├── assets/ # MJCF model files
│ │ ├── envs/ # Gymnasium environments
│ │ ├── scripts/ # Training & utility scripts
│ │ ├── tests/ # Pytest test suite
│ │ └── README.md
│ ├── brachiosaurus/ # Brachiosaurus (quadrupedal sauropod)
│ │ ├── assets/ # MJCF model files
│ │ ├── envs/ # Gymnasium environments
│ │ ├── scripts/ # Training & utility scripts
│ │ ├── tests/ # Pytest test suite
│ │ └── README.md
│ ├── trex/ # T-Rex (large bipedal predator)
│ │ ├── assets/ # MJCF model files
│ │ ├── envs/ # Gymnasium environments
│ │ ├── scripts/ # Training & utility scripts
│ │ └── tests/ # Pytest test suite
│ └── shared/ # Shared base classes and utilities
│ ├── base_env.py # BaseDinoEnv abstract class
│ ├── config.py # TOML configuration loading
│ ├── curriculum.py # Curriculum learning manager
│ ├── metrics.py # Training metrics tracking
│ └── wandb_integration.py
├── configs/ # TOML hyperparameter configs per species/stage
├── notebooks/ # Jupyter notebooks for experiments
│ ├── velociraptor_training.ipynb
│ ├── brachiosaurus_training.ipynb
│ ├── trex_training.ipynb
│ └── jax_trex_training.ipynb
├── website/ # Documentation site (Docusaurus)
└── Images/ # Training visualizations
Status: Active development
A bipedal predator with distinctive sickle claws, trained using 3-stage curriculum learning:
- Balance - Learn to stand without falling
- Locomotion - Walk and run forward
- Strike - Sprint and attack prey with claws
| Feature | Details |
|---|---|
| Observation | 69 dims (joints, pelvis, prey tracking) |
| Action | 12 dims (leg + claw controls) |
| Model | environments/velociraptor/assets/raptor.xml |
Status: Active development
A quadrupedal sauropod herbivore with a long neck for reaching elevated food sources. The first quadrupedal species in the project, featuring columnar elephant-like legs and characteristic longer front legs.
Trained using 3-stage curriculum learning:
- Balance - Stable quadrupedal stance
- Locomotion - Coordinated four-legged walking
- Food Reach - Walk to food and reach with neck
| Feature | Details |
|---|---|
| Observation | 75 dims (joints, torso, food tracking) |
| Action | 22 dims (6 neck + 16 leg controls) |
| Model | environments/brachiosaurus/assets/brachiosaurus.xml |
Status: Active development
Large bipedal predator with a massive skull, powerful jaws, and vestigial forelimbs. Hunts by sprinting toward prey and delivering a bite.
Trained using 3-stage curriculum learning:
- Balance - Stable bipedal stance
- Locomotion - Walk and run toward prey
- Hunting - Sprint and bite prey with jaws
| Feature | Details |
|---|---|
| Observation | 77 dims (joints, pelvis, prey tracking) |
| Action | 14 dims (3 neck/head + 1 jaw + 5 per leg) |
| Model | environments/trex/assets/trex.xml |
- Deinonychus (pack hunter)
- Allosaurus (large theropod)
- Compsognathus (small, fast biped)
# Clone and setup
git clone https://github.com/kuds/mesozoic-labs.git
cd mesozoic-labs
python -m venv venv
source venv/bin/activate
# Install the package with training dependencies
pip install -e ".[train]"
# View the velociraptor model
python environments/velociraptor/scripts/view_model.py
# Train stage 1 (balance)
cd environments/velociraptor
python scripts/train_sb3.py train --stage 1 --timesteps 1000000Hardware: Google Colab T4 GPU
| Dinosaur | Algorithm | Avg Reward | Training Time | Steps |
|---|---|---|---|---|
| Basic Dinosaur | PPO | 319.94 | 1:29:43 | 2,600,000 |
| Basic Dinosaur | SAC | 3091.31 | 4:36:59 | 3,600,000 |
| T-Rex | PPO | - | - | 5,000,000 |
| T-Rex | SAC | - | - | 5,000,000 |
| Velociraptor | PPO | 118.37 | 3:38:47 | 6,000,000 |
| Velociraptor | SAC | - | - | 5,000,000 |
| Brachiosaurus | PPO | - | - | 3,500,000 |
| Notebook | Description |
|---|---|
notebooks/velociraptor_training.ipynb |
Velociraptor 3-stage curriculum training (Colab-ready) |
notebooks/brachiosaurus_training.ipynb |
Brachiosaurus 3-stage curriculum training (Colab-ready) |
notebooks/trex_training.ipynb |
T-Rex 3-stage curriculum training (Colab-ready) |
notebooks/jax_trex_training.ipynb |
JAX/MJX T-Rex training for TPU acceleration (Colab-ready) |
- Complete velociraptor 3-stage training
- Complete brachiosaurus 3-stage training
- T-Rex environment and training
- JAX/MJX migration for faster training
- Multi-agent pack hunting scenarios
- Terrain adaptation (uneven ground, obstacles)
- Sim-to-real transfer experiments
See ROADMAP.md for the full phased timeline, milestones, and dependency graph.
- Documentation: mesozoiclabs.com
- Blog: From Zero to Dino-Roar
Contributions welcome! Open an issue or PR.
MIT License