Self-driving cars in the browser

First, we tried Craig Reynolds' flocking system. That's a repulsive field approach. It just won't work in a narrow space. We tried a system where there's a target projected in front of the vehicle, and repulsion fields from the object act on that. The vehicle then chases the target. This was a little better, but not much.

We ended up using a system where we computed a clear path ahead of the vehicle composed of two arcs, which could have opposite curvature. We always kept enough clear path ahead for the stopping distance of the vehicle, or we'd slow down. Arcs were tested against the map by brute force, then tweaked a bit for maximum obstacle clearance. The minimum turning radius increased with speed and decreased with vehicle tilt (this was for off-road operation), so the system wouldn't generate a path that would roll the vehicle.

(Our biggest headache at the event wasn't obstacle avoidance, it was that we took DARPA's prescribed GPS boundary limits as hard boundaries. The GPS positions were not accurate enough for that. After a very frustrating first run, where we could barely make it through narrow gates because the GPS boundaries and the real world obstacles were misaligned, we allowed 1 meter outside the boundary limits, and performance improved.)

The reversing is undeveloped and collision-detection is too basic (it looks like you're using a really small number of sensors). A car can get wedged between an angular object easily (and most times does).

As well, it seems like cars travel a set path and if gets obstructed, they get stuck trying to ram through it.

More explanations here http://lab.janhuenermann.de/article/learning-to-drive

I only say this because as I was reading it, I was looking at it from that standpoint, because on initial review, with the 2D nature and the "cartoonish" graphics (nothing wrong with that - in fact, it makes it very approachable), there is a subtle expectation that this demo might be for beginners in the field.

I continued reading, and as I read the technical background, it became quickly apparent that this was not using a simple multi-layer backprop ANN, but something seemingly more advanced, with little to go on to understand how it worked. There is enough there that an interested learner could research the topic, but I think a beginner to the topic would quickly become overwhelmed, because none of the various jargon and other terms ("hyper-parameters? what the hell are those?", I can hear such a newbie exclaiming) have been defined.

I don't believe you ever intended this to be understood or used as an introduction or a beginner's tutorial to machine learning, but because it appears like it could be, mentioning that it isn't at the beginning might be appropriate.

Furthermore, this could be an opportunity for you or someone else to create an approachable series of interactive "run-in-the-browser" javascript machine learning tutorials, starting with the basics, and culminating in something like this. Such a series would need to be planned well, and first defined what should be shown and in what order (ie - Do you discuss basics of linear algebra or probability? Do you incorporate classic ML techniques in the series along with neural networks?). Hmm - now that I think about it, I can actually envision a whole sequence of series of lessons, building from the bottom ("What is a vector?") on up...

Ultimately - I found this demo fascinating and interesting. Thank you for sharing it!

I don't think it really approaches self driving cars though until there are lanes, stop signs, and other road obstacles.

Edit: They crashed into each other head first. They eventually sorted it out.

Reminds me of a child hood car I had which would bump against a wall, climb it slightly and then use the gravity to drop down and turn.

http://pasteboard.co/DpvfypZ6B.png

I think I found a small bug: cars don't seem to have much ability to estimate the size of a gap between obstacles, so they'll attempt to squeeze through gaps that are too small for them.

I just have to ask if someone knows: Is there any benefit to doing this in WebGL? Does that enable you to use the GPU for other calculations than the graphical ones? (I really don't know anything about GPUs, so sorry if that's a naive question).

Surprisingly, they seem to successfully avoid user-added obstacles.

Also, the cars need to learn to drive on one side of the road.

Also, it would be nice if people could upload their own javascript to the environment (i.e., like JSFiddle but with car-simulating capabilities).