Back to the topic of neural networks. I contacted Amir Efrati at The Information, but I have not been able to get to the bottom of this. Is Tesla working on a neural network for path planning? If so, is Tesla using some form of imitation learning to train its path planning neural network?
The full scoop: Tesla AI and behaviour cloning: what’s really happening?
A most curious and puzzling case...
There is a big difference between what Tesla does, and what Tesla imagines is possible.
They talk about detecting objects to steer around, but I'm not aware of any generic "blob" object detection.
In experiencing how it drives I don't see any signs that there is any behavior cloning going on. I really wish there was, and that I could train it on my own driving. Having autosteer, and auto acceleration braking that was set to my own behaviors would be pretty cool even if it took months of shadow training. It would be awesome if slight adjustments in steering helped correct it.
The biggest challenge facing the AP team likely has to do with the fact that it lacks important information to make sense of why I did what I did while driving.
It often has no clue about the speed limit on the road
It doesn't detect debris in the road
It doesn't have any kind of rear/side radar to overcome imaging limitations
It says it has drive-on-nav, but this doesn't prevent it from diving for an exit that isn't the exit in the nav
It has no ability to measure my engagement (the torque sensor is really inadequate for this). So it doesn't know how to interpret my allowance of whatever it is doing. If I ignore a lane change request I could have a bunch of reasons for doing so. If I take over it might be because I had to or because I simply wanted to.
If there was behavior cloning then I'd like there to be a way to disable it temporarily. This falls under the do what I say, but not what I do column of driving.
I'm sure we all have driving moments where we'd scrub from the memory banks.
