Skip to main content

Welcome to Mesozoic Labs

Mesozoic Labs is an open-source platform for building robotic dinosaurs using physics simulation and reinforcement learning.

What is Mesozoic Labs?

We're bringing prehistoric creatures back to life—as robots. Using state-of-the-art reinforcement learning algorithms and accurate physics simulations, we train virtual dinosaurs to walk, run, and move naturally. Then we transfer that learned behavior to real robotic hardware.

Key Features

  • Accurate Physics Models - MuJoCo-based simulations of dinosaur anatomy
  • Reinforcement Learning - PPO and SAC algorithms for training locomotion
  • JAX/MJX Integration - GPU-accelerated training with 2,048+ parallel environments for 10-100x speedup
  • Multiple Species - T-Rex, Velociraptor, Brachiosaurus, and more
  • Sim-to-Real - Transfer learned behaviors to physical robots
  • Open Source - Fully open codebase for research and education

Current Results

SpeciesAlgorithmStageBest Eval RewardAvg Fwd VelTraining StepsTime
T-RexPPO1 — Balance3008.66 +/- 7.620.02 m/s6.0M3h 35m
T-RexPPO2 — Locomotion1936.01 +/- 13.123.47 m/s8.0M5h 17m
T-RexPPO3 — Bite1294.28 +/- 67.191.68 m/s8.0M4h 10m
VelociraptorPPO1 — Balance1964.43 +/- 27.390.11 m/s6.0M2h 57m
VelociraptorPPO2 — Locomotion2678.68 +/- 4.073.47 m/s8.0M4h 36m
VelociraptorPPO3 — Strike1366.19 +/- 76.292.02 m/s8.0M3h 52m
Basic DinosaurPPOFull curriculum319.942.6M1h 30m
Basic DinosaurSACFull curriculum3091.313.6M

All results use 3-stage curriculum learning (balance → locomotion → behavior).

Project Status

The project is actively under development (currently on Phase 1 — v0.3.0). Core infrastructure is in place including automated curriculum training, W&B experiment tracking, and evaluation metrics. Star us on GitHub to follow our progress!