Sim-to-Real: Simulators & Transfer
Training robots in virtual physics engines. Traces how policies learn in sim and successfully transfer to real-world steel.
Simulation is cheap and runs 100x faster than real-time, but mismatches in friction, sensor noise, and gear play can cause sim-trained models to fail instantly in reality.
Physics simulators: MuJoCo & Isaac Sim
Because real-world training is wear-heavy and slow, labs train policies in virtual physics simulators like MuJoCo or Nvidia's Isaac Sim. These simulators model rigid body dynamics, joint torque limits, contact forces, and cameras, running millions of parallel episodes in minutes.
The robot learns to walk, balance, or grab cups entirely in code before it ever touches physical steel.
Domain Randomization: training for chaos
Simulators are approximations; they cannot capture real-world complexity perfectly (the sim-to-real gap). To prevent the policy from overfitting to simulator defaults, labs use domain randomization.
During training, the simulator constantly randomizes joint mass, link lengths, motor delays, floor friction, lighting, and camera noise. By forcing the neural network to succeed across all these variations, the policy learns to be robust enough to handle the real world.
System Identification: calibrating the virtual world
System identification is the inverse of domain randomization. Engineers run calibration routines on the physical robot, measuring actual joint velocities, temperatures, and gear backlash. These real-world telemetry parameters are fed back into the simulator to tighten the physics constants, narrowing the sim-to-real gap.