Skip to main content

Welcome to Mesozoic Labs

Mesozoic Labs is an open-source research platform for dinosaur-inspired locomotion using physics simulation and reinforcement learning.

What is Mesozoic Labs?​

We use articulated MuJoCo models to study how reinforcement-learning agents can learn balance, locomotion, and species-inspired tasks. These models are engineering abstractions inspired by dinosaur morphology; they have not been validated as accurate reconstructions of dinosaur anatomy. Hardware transfer is a future research direction, and no sim-to-real or physical-robot results are currently published.

Key Features​

  • Physics-Based Models - Articulated MuJoCo models inspired by dinosaur morphology
  • Reinforcement Learning - PPO and SAC algorithms for training locomotion
  • Behavior Recipes - Every model trains stand, walk and hunt (and recover, on the models with a recovery stage) as separately certified, separately published policies on one shared trunk, declared in configs/<species>/<embodiment>/stages.toml
  • Six Models of Five Species - The anatomical T-Rex, Velociraptor, Brachiosaurus, Dibothrosuchus and Compsognathus, and the Compsognathus robot; each model is a species and an embodiment (anatomical or robot)
  • Sim-to-Real Roadmap - Planned hardware, system-identification, and transfer experiments; not yet validated
  • Open Source - Fully open codebase for research and education

Species Catalog and Published Results​

The model pages render observation and action dimensions, compiled-model facts, current curriculum stages, and published result summaries from the generated species catalog. Each stage row shows its recipe label (stand, recover, walk or hunt), whether it is a published deliverable, and the parent it warm-starts from; published results list one row per deliverable, headlined by the metric its gate kind certifies:

The published result summaries describe historical ladder runs with no per-deliverable records. They are marked unverified because the original repository commit, model hash, and config hash were not recorded; they should not be treated as controlled algorithm comparisons or as results from the current model revision and configs.

Project Status​

The project is actively under development. Core infrastructure includes automated curriculum training, W&B experiment tracking, and evaluation metrics. See the repository roadmap for planned work, including hardware and sim-to-real experiments that have not yet been validated.