Yea they could put down visual markers on the road to make it easier for CV or ML models to be trained against, or maybe even physical tracks to mechanically guide these cars so that minimal software is needed. Multiple cars can be linked together for efficiency.
Right. But that is in designing the existing roads. I'm talking about the cars. And I'm specifically asking why not adding more options? I don't mind a camera being part of the solution at all. Gives an obvious path to human labeling of training data. But, why not have more?
Those cars generate a lot of sensor data. (Tb per drive?). Id imagine that data needs to be made actionable and seperated into training and simulation sets pretty quickly. Mapping is a massive problem to automate.
I imagine they'll just have to do what humans do. Follow existing tire tracks if available. If not, just intuit where the road using your best judgement. It's not terribly unlike driving on dirt roads (though harder, surely) which autonomous cars already did surprisingly well in the first two Grand Challenges (no lane markers there either).
Yes, according to Karpathy. The cars upload data (typically still images) which are then labeled by humans and used as training data to improve the models.
Most teams use ml heavily for perception but not necessarily for other things. You need explainability when it comes to safety, and maps are needed no matter what if you want your car to get around.
I recall seeing that option discussed on The Daily Planet with Jay Ingram back in the early 00s. They interviewed a research engineer on self-driving vehicles, probably in the context of the 2004 DARPA Grand Challenge. He'd clearly been working in the space for years and was very knowledgeable.
IIRC, they had some very successful prototype road systems back in the 90s (or maybe earlier) that were based on implanting magnetic markers into the road at regular intervals. The big problem was cost. There's miles and miles of roads and the cost to upgrade and maintain enough miles was prohibitive.
With that said, perhaps now that the cars themselves are 90% of the way there, the upgrades to the roadways could be more limited. I wonder if adding machine-readable markers to particularly important or problematic areas might be reasonable at this point.
I'm sure this isn't what the OP had in mind but I think it would help adoption. At first you see one or two of the cars trickle onto the roadways and you see them behave. Then a few more. I think this would help the comfort level of people concerned about the safety and efficacy of AI driving.
One possibility, is that if they can get good enough inital mapping data, and good enough self-driving cars, the cars could update the mapping data while they're working.
That's a lot of trust on the self-driving cars, though.
When all the cars on the road are self driving, I don't see why they would need to have all their radars on at once. I imagine they would be able to communicate with each other and form a complete picture of the entire roadway without all of them having their radars on (or perhaps running at a lower power?)
This would be very difficult to train with ML, however. I'm sure there are other downsides I'm not thinking of
I'd imagine this could work on a small fleet of miniature cars (rc size?). A model town/city could be built with various obstacles for training a large amount of the ai, I would think.
I'm thinking the same-thing but they could even take this data out of the self driving car, apply it to a normal car and even in that case it would be a huge change that could be incredibly beneficial!
reply