Physics Simulation

Watch It Walk Before You Build It

MuJoCo physics simulation of humanoid locomotion. The same engine used by Google DeepMind, ETH Zurich, and every serious robotics lab. These simulations use real physics — gravity, contact, friction, motor torque limits — not animation.

Unitree G1 — MuJoCo Walking Simulation

Trained with RSL-RL (ETH Zurich's reinforcement learning framework) in Isaac Lab. The policy learns to walk from scratch in ~36 hours of simulated experience across 4,096 parallel environments. No motion capture. No hand-tuned gaits. Pure RL.

MuJoCo 3.xRSL-RL (PPO)4,096 parallel envsZero-shot sim-to-real

What This Means for Kit Builders

Validate Before You Build

Run your kit's BOM in simulation before ordering $10K+ of parts. Does the hip actuator have enough torque? Does the battery last? Does it fall over? Simulation answers these in hours, not months.

The Software Stack

MuJoCo (physics) + RSL-RL (RL algorithm) + Isaac Lab (parallel training) + ROS2 (deployment). This is the same stack Tesla, Unitree, and Agility use. See our Software Stacks page for the full breakdown.

From Simulation to Kit

Our compiler validates your BOM across 9 physical dimensions. Simulation validates it actually walks. Together, they close the gap between “compatible parts list” and “working humanoid.”

The Sim-to-Real Pipeline

1

Define Robot (URDF/MJCF)

Joint layout, mass, actuator limits from your kit BOM

See our H1 URDF →
2

Train in Simulation (Isaac Lab + RSL-RL)

4,096 parallel environments, 36 hours, PPO algorithm

Software stack details →
3

Adaptive Domain Randomization

Auto-vary mass, friction, motor gains, terrain — builds robustness

4

Validate (Humanoid-Gym benchmarks)

Zero-shot sim-to-real metrics before touching hardware

5

Deploy to Real Hardware (ROS2)

Trained policy runs as a ROS2 node on your kit's Jetson

See kit options →
6

Brief Real-World Fine-Tuning

1-4 hours on real hardware to close the sim-to-real gap

Live Browser Simulation — Coming Soon

MuJoCo compiled to WebAssembly, running real physics in your browser. Pause, scrub, rotate, swap robots. No install required. We're building this using mujoco_wasmwith the H1 MJCF model from Google DeepMind's MuJoCo Menagerie.

MuJoCo WASM: ~4 MB·H1 meshes: ~9 MB·Target: 40-60 FPS on desktop

Resources

Ready to build a walking humanoid?

Start with the compiler to validate your BOM, then use simulation to validate locomotion.