Science1 publisher2 min readPublished
Stanford world model docks with the ISS about 53% of the time in preprint tests
Stanford's Out-of-this-World-Model docked with the ISS about 53% of the time in tests, after 50 times fewer training runs than a comparable learning system. The cheaper training helps researchers, though a failure rate near 47% leaves flight certification far off.
The Scientist · Science desk
What happened
- Stanford researchers posted a preprint to arXiv describing the Out-of-this-World-Model, an AI that runs learned 'mental simulations' of a spacecraft's approach in order to dock.
- Training needed hundreds of thousands of simulated flights, weeks of work on CPUs, so the team wrote a GPU library called AstroJAX to do it.
- The model needed 500,000 training iterations to learn docking, against 25,000,000 for a comparable reinforcement-learning system.
- Across all of the International Space Station's docking ports, the model docked successfully about 53% of the time.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- cost A 50-fold cut in training runs lowers the GPU time each round of research on learned docking controllers costs, so labs can try more designs for the same budget.
- constraint With roughly 47% of attempts failing, the controller is far from use on a crewed approach, where one collision could kill everyone aboard both vehicles.
- contradiction phys.org says the model needed 500,000 iterations to 'master' docking, yet it succeeded about 53% of the time, so the 50-fold figure counts runs to a training milestone and says little about reliability.
Divide the baseline's 25,000,000 training runs by the Stanford model's 500,000 and the ratio is 50 [1]. The figure counts simulated practice before each model could dock. The phys.org account gives the Stanford number in "iterations" and the baseline's in "permutations" [9], so the comparison is only as clean as those two units are alike. The saving is paid once, on GPUs, during training [8]. A flight computer would pay on every decision, because the model works by imagining dozens of possible futures and steering the spacecraft toward one of them [6].
The design responds to specific failures in the existing tools. Conventional navigation pairs a guidance, navigation and control algorithm with an extended Kalman filter that turns GPS and star-tracker readings into thruster burn duty cycles [3]. Kalman filters handle high-speed video poorly. The computer-vision methods used in their place can be thrown off by sunlight glinting from a solar array or a sudden shadow on the target [4]. Reinforcement learning, the next thing researchers tried, works only under a fixed set of rules. Move the docking port to the other side of the station and the algorithm no longer knows what to do [5].
A moved port is where reinforcement learning broke, so the unfamiliar-port test is the one I would look at first. The OWM did better at a port it had never seen, and coped better when the researchers placed a docked capsule on the port it was meant to use, according to phys.org [10]. That result makes it unlikely the model had simply memorised one approach path.
Certification is the harder question. Engineers have docked for decades with hand-written physics equations and human pilots making corrections [2]. Equations someone wrote can be inspected term by term. A learned model's physics is whatever it picked up from experience in training [6]. The OWM's partial answer is to attach a probability to each predicted outcome, and phys.org calls that confidence measure key to adjusting when predictions do not come true [7].
I think that probability is the most useful feature here for whoever eventually has to sign off on a learned docking controller, on one condition. A model reporting high confidence has to succeed about as often as it says it will. The account of the preprint does not report calibration results for this one.
What to watch
- A revised version of the preprint, or a peer-reviewed paper, that reports tests on hardware or in orbit.
- Calibration data showing whether the OWM's stated outcome probabilities match how often it actually docks.
- Whether further training moves the success rate across all ISS ports well above 53%.