Trying to do self-driving entirely by deep learning in a reactive system seems like a terrible idea. Deep learning to classify objects (bicycles, pedestrians, cars, traffic lights, telephone poles, cops) is fine. But map building, planning, and obstacle avoidance needs to be more reliable than a purely reactive system can do.
Look at the videos from Urmson's talk at SXSW. That shows the worldview of a Google self-driving car. It's about 80% geometry and 20% classification.
Yes, you can get a pure deep learning system to drive on a freeway. But how does it do in a more cluttered environment? A system that builds local maps and profiles terrain with LIDAR can deal with clutter.
Why exactly can't you use deep learning to drive in a "cluttered environment?" You didn't mention why it's a terrible idea (I have no background in deep learning, just curious)
Totally agree on this. Making neural networks build a debuggable scene graph is a pretty hard challenge. With the scene graph things become quite easy.
Look at the videos from Urmson's talk at SXSW. That shows the worldview of a Google self-driving car. It's about 80% geometry and 20% classification.
Yes, you can get a pure deep learning system to drive on a freeway. But how does it do in a more cluttered environment? A system that builds local maps and profiles terrain with LIDAR can deal with clutter.