← sources

RL agent learning to land a Falcon 9 on a droneship (fully simulated, 67 randomized parameters)

source: reddit (reddit:reinforcementlearning)published 2026-10-20 10:09ingested 2026-10-21 02:40

retained text

An RL agent learning to land a Falcon 9 on a droneship. Everything is simulated: rocket physics, waves, wind, fuel slosh, sensor noise. No real flight data, just trial and error.

It’s a hoverslam, so the booster can’t hover and has to hit zero speed right at the deck.

PPO, \~80k parameter policy, 10 controls (throttle, gimbal, grid fins, thrusters, legs, ignition), 67 world parameters randomized every attempt.

Out of 59M training flights: 34.7M splashed, 10.8M exploded, 13.2M landed.

Still weak in strong wind and high waves, working on that now. Happy to answer questions.

document fetched 2026-10-21 00:38 · canonical https://reddit.com/r/reinforcementlearning/comments/1xao7o3/rl_agent_learning_to_land_a_falcon_9_on_a