PHASE 0 — v0.2.0
Clean Slate
- TOML configs
- Gymnasium registration
- Developer tooling
- Testing improvements
MESOZOICLABS
PREHISTORICSIMULATEDINTELLIGENCE
Open-source platform for dinosaur-inspired locomotion research
powered by physics simulation and reinforcement learning
- SPECIES
- 6
- ACTUATORS
- 120
- PUBLISHED RUN STEPS
- 100M
Core Features
MuJoCo Physics
Physics-based articulated models inspired by dinosaur morphology, with contact dynamics and actuator models.
Reinforcement Learning
PPO and SAC algorithms via Stable-Baselines3 with automated 3-stage curriculum learning.
6 Species
Velociraptor Mongoliensis (22 actuators), Tyrannosaurus Rex (15 actuators), Brachiosaurus Altithorax (30 actuators), Dibothrosuchus Elaphros (27 actuators), Compsognathus Longipes (14 actuators), Compsognathus Longipes (Robot) (12 actuators).
Open Source
MIT licensed with TOML configs, Gymnasium registration and W&B tracking.
How It Works
Each species learns through a 3-stage curriculum, progressively mastering more complex behaviors
STAGE 1
Balance
Stand & Stabilize
The agent learns to stand upright and maintain balance using proprioceptive feedback and joint torques.
STAGE 2
Locomotion
Walk & Run
Building on balance skills, the agent develops forward locomotion with coordinated movement patterns.
STAGE 3
Behavior
Species-Specific Task
Species-specific simulator tasks add sickle-claw contact, a fixed head-contact proxy, or head-tip target proximity.
See It In Action
PPO Stage 1 — Balance
Published Run Summaries
Provenance-labelled metrics and the curriculum videos currently available in the repository
Sickle-claw contact attacks
- Task success
- 93%
- Max reported avg. velocity
- 3.47 m/s
- Timesteps
- 22M
- Run
- PPO · Stable-Baselines3 · 2026-03-15
historical model · unverified · evaluation episode count not recorded · Stable-Baselines3 version not recorded. Metrics are retained experiment records, not claims about the current model revision.
STAGE 1PPO · Stable-Baselines3 · historical
Balance
Learn to stand and balance without falling
unverified artifact; backend version not recorded; not evidence for the current model or config.
STAGE 2PPO · Stable-Baselines3 · historical
Locomotion
Learn forward walking/running
unverified artifact; backend version not recorded; not evidence for the current model or config.
STAGE 3PPO · Stable-Baselines3 · historical
Strike
Sprint and strike prey with sickle claw
unverified artifact; backend version not recorded; not evidence for the current model or config.
Project Roadmap
PHASE 1 — v0.3.0
First Steps
- Curriculum manager
- W&B tracking
- Locomotion metrics
- Species training runs
PHASE 2 — v0.4.0
Into the Wild
- Domain randomization — planned
- Terrain diversity
- Turning & steering
PHASE 3 — v0.5.0
Evolution
- Custom policy networks
- New species
- Benchmark suite
PHASE 4 — v0.6.0
The Pack
- Multi-agent envs
- Cooperative hunting
- Predator-prey
PHASE 6
Sim-to-Real
- Hardware prototype — planned
- ROS 2 bridge — planned
- Sim-to-real validation — not started
Ready to Train a Dinosaur-Inspired Agent?
Explore the docs, train your first model, or contribute to the project.