Stanford’s ‘Out-of-this-World-Model’ lets spacecraft mentally simulate docking maneuvers

# What the team built

# Why docking is hard Docking in Low Earth Orbit is a demanding control problem. Relative speeds are around 28,000 km/h, motion is constrained by orbital mechanics (accelerating in one direction can cause drift in another), and collisions have catastrophic consequences. Traditional Guidance, Navigation and Control systems combine model-based filters such as Extended Kalman Filters with engineered computer-vision pipelines. Those vision approaches break under lighting changes, glints, or partial occlusion.

# How a world model works here

# Training at scale with AstroJAX Teaching a world model to handle docking requires hundreds of thousands of simulated flights. To make that tractable, the researchers created AstroJAX, a GPU-accelerated dynamics library intended to run on the same hardware used for modern neural-network training. Running the simulations on GPUs reduced training time and enabled OWM to learn with far fewer iterations than the reinforcement-learning baseline.

# Performance and limitations In head-to-head comparisons, OWM required about 500,000 training iterations to learn docking maneuvers, while a comparable reinforcement-learning system needed roughly 25,000,000. OWM achieved a success rate of about 53% when tested across docking ports on the ISS, compared with 29% for the RL approach. OWM also generalized better to previously unseen docking port locations and tolerated certain unexpected disturbances—one test introduced an already-docked capsule at the target port.

However, OWM had trouble with close-range operations. The authors attribute this partly to heavy penalty weighting for collisions during training, which pushed the model to be overly conservative at close distances. The paper indicates more tuning and development are needed before deployment in missions that carry human passengers.

# Practical implications and next steps

# Bottom line