How much AI/ML do you think is running on the software associated with this launch and docking? I suspect there’s a great deal of very explicit, well tested code, but very little AI/ML.
Very little in terms of what you would consider AI or ML. It's a lot of trig and control theory. You do a lot of the math beforehand to figure out the script the craft will follow to reach the ISS. Once it gets close, you either use manual controls or let autopilot do the docking. The autopilot is basically looking some dots on a camera and firing the thrusters to make the dots line up right. It's more like a dumb PID controller than an AI. The autopilot isn't learning anything or somehow making connections that weren't already programmed. This is all stuff NASA figured out how to do in the 60s so it's nothing novel except for the mission and the hardware the math is done on.
Dragon is using a flash LIDAR made by Advanced Scientific Concepts.[1][2] 128x128 pixels.
The Russians used a KURS unit, which is a cooperative (the ISS has a transponder to talk to it) radar-based system. That was from Ukraine, and after going to war with Ukraine, Russia had to develop an in-house system.
All of this stuff is older technology, predating machine learning.
Just like regular autopilot on the ground, the AI/ML is really just in the image / signal recognition. After that it’s mostly control law (rules based system, and force calculations0.
What you wrote is what I would have guessed. Are you also guessing or do you know for sure? It sounds right, but I'm very curious about the actual answer to this question.
I don't work in the industry, but I have the training, and can confirm what he said. No AI or ML. We use a combination of closed-form solutions and deterministic numerical analyses. There is no actual need for AI or ML in this problem domain, outside of perhaps the design phase. Sometimes dynamic programming and some ML techniques are used when, I dunno, figuring out the best configuration for an antenna or profile of a nozzle. But flight dynamics and control do not need AI, as the physics are well-understood.
