https://blog.nelhage.com/2011/03/exploiting-pickle/

> The biggest problem I ran into was over fitting the model so that it would not work in evenlly slightly different scenarios.

Regardless, a very cool project.

This simple project is a good example of supervised learning from what I can tell - the network will learn to steer "as good as" the human that provides the training data. For a different (and more complex) flavor of algorithm, check out reinforcement learning, where the "agent" (computer system) can actually learn to outperform humans. Stanford's autonomous helicopters always come to mind - http://heli.stanford.edu/

- For liability reasons, most of the algorithmic IP will likely be open sourced. Either because it's required by regulators or because it's the most efficient way for car makers to socialize risk of an algorithmic failure.

- Electric vehicles have many fewer moving parts, which means that the remaining parts are likely to be converged upon by the industry and used widely. This breaks a lot of platform-dependency issues and allows for the commoditization of parts like motors. As these become standardized and commoditized, and easily comparable on the basis of size, torque, and efficiency, there will be virtually no benefit to carmakers to manufacture their own. The same applies to aluminum monocoque frames, charging circuitry, etc.

Tesla currently differentiates its models based on how many motors and what size batteries, but beyond that it's mostly just cabin shape, along with new innovations like the hepa filter cabin air cleansing which will likely be a standard part of all future models.

- Battery tech works the same way as motors, with little competitive advantage to be gained by automakers, especially since most of the IP in this area is already spoken for.

Compare the number of patentable parts in a model T vs a 1998 Taurus vs a 2017 internal combustion vehicle vs a Telsa. Tesla is one innovator, and GM has already likely patented many inventions relating to EV technology back in the original Chevy Volt era.

All this is why Tesla acquired SolarCity and is attempting to make an infrastructure play rather than a technology play. Only due to Musk's rare ability to self-finance big risks is this even possible, since infrastructure moonshots featuring $30K+ hardware units are hard to fund.

I think back to the days when I basically implemented lane following with an array of photo resistors, an Arduino, a shitty robot made from Vex parts and some c code. The problem is much simpler than the one presented in this article, but then the computational resource used was order of magnitudes less. At what point then, do you decide that "OK I think the complexity and nature of the problem warrants the use of ML" or "Hmmm I think neural network is an overkill here"?

Thanks!

https://github.com/samjabrahams/tensorflow-on-raspberry-pi

Note: I am the owner of this repo

Using Keras' "validation_split" parameter will just randomly select a validation set. This is not the right thing to do when your data is image sequences, because you will get essentially identical data in training and validation.

Because of this, the numbers/plot here might as well be training accuracy numbers.

But rather than a black box, it's explainable what the different layers are doing. If neural nets are turing machines then we should be able to compile some parts of the net from code.

Then the net is a library of layers. Some Layers trained with back prop, some compiled from code.

Awesome tutorial.

It uses a Raspberry Pi and ~50 lines of code. So I don't think anyone should expect it to do something that's impossible with other approaches.

Is it really trivial? Honest question... Which control algorithms are you speaking of?

While this example is simplified - and I wouldn't recommend it for a real-world full-size vehicle trial - it does implement everything (scaled down) described in NVidia's paper:

https://images.nvidia.com/content/tegra/automotive/images/20...

In short, the project uses OpenCV for the vision aspect, a small CNN for the model, uses "behavioral cloning" (where the driver drives the vehicle, taking images of the "road" and other sensor data like steering - as features and labels respectively - then trains on that data), and augmentation of the data to add more training examples, plus training data for "off course" correction examples...

If you read the NVidia paper, you'll find that's virtually all the same things they did, too! Now - they gathered a butt-ton (that's a technical measurement) more data, and their CNN was bigger and more complex (and probably couldn't be trained in reasonable time without a GPU), plus they used multiple cameras (to simulate the "off-lane" modes), and they gathered other label data (not just steering, but throttle, braking, and other bits)...but ultimately, the author of the smaller system captured everything.

Furthermore, NVidia's system was used on a real-world car, and performed quite well; there are videos out there of it in action.

This is virtually the same kind of example system that the "behavioral cloning" lab of Udacity's Self-Driving Car Engineer Nanodegree is using. We're free to select what and how to implement things, of course, but I am pretty certain we all understand that this form of system works fairly well in a real-world situation, and so most of us are going down the same route (ie, behavioral cloning, cnn, opencv, etc). Our "car" though is a simulation vehicle on a track, built using Unity3D.

> Of course, the CV part -- "seeing" the lines -- requires some form of ML to work in the real world.

Actually, it doesn't. The first lab we did in the Udacity course used OpenCV and Numpy exclusively to "find and highlight lane-lines" (key part was to convert the image from BGR to HSV, and mask using the hue). No ML was required.

That said - I wouldn't trust it for real-world vehicle driving use - but it possibly could be used as part of a system; however, as NVidia has shown, a CNN works much better, without needing to do any pre-processing with OpenCV to extract features of the image - the CNN learns to do this on its own.