It sounds like what they have is machine vision as a novel input into a conventional control loop. For one panel at a time, wiggle it about and look how the light spot changes. Find the local maximum for tightest spot. That's now "locked" in the PLL sense and should stay locked for days or weeks.
I'm sure there are tons of considerations that aren't occurring to me (for instance, thermal expansion likely contributes substantially to making it a moving target (literally)), but I really don't see where "AI" comes into it...
The so-called "AI" here is really just a clever way to use closed-loop control so that the above factors don't matter much.
(Side note: A large amount of "AI" these days just amounts to changing open-loop mechanisms to closed-loop, often by using computer vision as the state sensor. 3D printers, for example could become much more precise with no mechanical improvements if they did this. Probably some have already done so.)
I'm not sure how to make such a simple system for mirrors pointing at targets, but I suspect "AI" and "cameras" are overkill.
That's how it's been done for like 40 years. No vision or "AI" required.
Ground vibration might also be a problem, in applications where this is being used to generate heat for some heavy industrial application. Presumably in that cast said heavy industrial application would be taking place nearby, and so could disturb the ground under the mirrors or tower.
Control loops are everywhere, but CD-ROM drives do not use machine vision. That's why op said that's the novel part.
You also need precise surveying of the ground site to know its geometry. The precise actuators and surverying are expensive.
Their approach is using cheap actuators in a closed-loop system where a computer vision controller (using the four cameras on top of the tower) moves the mirrors until it gets it right. Cheap actuator slip and lag is terrible, but with the feedback of the closed-loop, it's doable.
This is an approach that's now allowing hobbyists to build cheaper motorised telescopes, robot arms and so on.
I suspect that the "AI" they mention is mostly photogrammetry - being able to tell the position and orientation of your mirrors from a video stream - in a feedback loop with motor controllers and temperature sensors. Use of cameras would make it close enough to "AI" that the buzzword would stick.
This. Modern photogrammetry uses computer vision techniques like deep neural nets.
It's an amazing advancement.
Imagine this problem: having to place a ground vehicle like a wheeled robot at an arbitrary coordinate in a room. The mirror positioning problem can be reduced to this one. Traditionally, you'd do it by building a very nice wheeled robot with non-slip wheels, wheel rotation angle sensors, fraction of a degree stepper motors, all very nice to work with, precise, and expensive. You could even solve this in an open-loop system with good enough actuators.
Today, just stick a camera on the ceiling - or on the robot, if want a more challenging problem - and use CV to move a cheap RC car where you want.
The Terminator will jangle and jiggle, instead of being built like a precise Swiss watch. But you'll get ten of them for the same price!
I'm surprised to hear that. What good do DNN do for photogrammetry? Aren't they too unreliable?
I only have university course-level exposure to this, most of it spent learning how to compute coordinate systems and transformation matrices from points on a set of pictures taken from different position. I guess DNNs could help with object isolation and labeling (identifying same objects on a set of pictures taken from different perspective)?
* https://www.youtube.com/watch?v=Ccj1O7yndIk
* https://www.youtube.com/watch?v=eOuonMhEsxI
* https://www.groundai.com/project/df-slam-a-deep-learning-enh...
The Related work section in the DF-SLAM paper is great.
As you intuited, DNNs work great for tracking visual landmarks (better than traditional computed features) between frames. They can also estimate depth from a monocular camera, which is very useful.
A more out-of-left-field idea is training a CNN to directly regress the camera pose (position and orientation) from an image:
The light is being reflected and focused by mirrors. That light is being directly onto some volume of material. So, in this case, they want to get the material above some critical temperature (e.g. so it melts). I imagine that uniformly heating the whole volume may not be the best approach, or that you may be able to work with smaller volumes depending on your expected solar input etc.
So I can see a little more justification here than the case where solar cells are just being pointed at the sun.