Trained with reinforcement learning

AI pilots that land rocket boosters

Neural networks learned to land real rocket boosters, and to catch one with the tower’s arms, by trial and error in a physics simulation.

What the simulation includes
  • Wind gusts
  • Wind shear
  • Wind turning with height
  • Turbulence
  • Updrafts
  • Microbursts
  • Rain
  • Thinning air
  • Waves
  • Ocean swell
  • A rolling droneship
  • A heaving droneship
  • A drifting droneship
  • A ship off its station
  • A wet, slippery deck
  • Sloshing fuel
  • An off-centre mass
  • Drag
  • Landing legs
  • Legs that crush
  • Legs that take time to deploy
  • Legs torn off by fast air
  • Sensor noise
  • Sensor delay
  • Engine lag
  • Ignition delay
  • Limited relights
  • Failed relights
  • Engines cutting out
  • Engine misalignment
  • Thrust jitter
  • Engine swivel limits
  • Three rings of Raptors
  • Fuel settling before a relight
  • Grid fins
  • Grid fin stall
  • Cold-gas thrusters
  • Limited thruster gas
  • GPS dropouts
  • Gyro drift
  • Accelerometer drift
  • Noisy actuators
  • Tower arms
  • A moving carriage
  • Shock absorbers on the rails
  • Bumpers on the arms
  • Wind shadow behind the tower
  • The tower’s own tracking
  • Catches called off