three ways to close the reality gap
the reality gap
sim physics approx = real physics + epsilon (friction, delay, lighting, noise, ...)
domain randomization
theta_sim ~ P(theta)
train over many sims
policy sees random gravity,
mass, friction, motor gain
zero real data needed
OpenAI Dactyl (rubik's cube)
over-rand. -> cautious
under-rand. -> brittle
system identification
measure real:
friction, delays, gains
plug into sim
train on matched sim
low noise, precise
but residual error invisible
to the policy
teacher / student
teacher: privileged info
terrain, ground truth
student: only sensor obs
distill teacher -> student
ANYmal, Boston Dynamics
robust across physical
parameters + obs noise
2026 recipe: GPU-parallel sim + DR + teacher/student + safety shield on real hardware