Engineers teach spacecraft to 'dream' their way to the space station

A new AI system developed by Stanford researchers enables spacecraft to "dream" their way to the space station, using a world model to simulate potential futures and control spacecraft movements. The system, called the Out-of-this-World-Model (OWM), uses reinforcement learning to navigate in low Earth orbit, avoiding the limitations of traditional Kalman filters. The OWM model requires only 500,000 iterations to master docking maneuvers, compared to a comparable reinforcement learning system that needed 25 million permutations. The model also performed better when docking at a never-before-seen port and handled unexpected situations, such as placing a docked capsule at the intended port. The model achieved a 53% success rate in docking across all ISS docking ports, though this should be compared with the RL algorithm's 29% success rate. The researchers used a GPU-accelerated library called AstroJAX to run the model, which is designed for efficient computation. The OWM model is still in development and may take time before it can fully automate docking sequences. The development highlights the potential of AI to improve spacecraft navigation and reduce reliance on traditional physics-based algorithms.

AI-generated summary of the original article — see the source for the full story.