Re:MIT 6.S094: Deep Learning for Self-Driving Cars
by bitL on Thursday January 18, 2018 @04:02AM
(#16175694)
AFAIK most SDC companies are using all kinds of control inputs, whether they are based on classical Computer Vision (RANSAC, HOG, SVM), on Deep Learning with 2D convolutions, on Deep Learning with 3D convolutions and RNN (2D + time on a series of images), Deep Reinforcement Learning (lane keeping, realtime path planning/obstacle avoidance), LiDAR+Radar+Ultrasonic combo via Kalman or particle filters, GPS, IMU etc. and then mix all those signals together to figure out which ones are the most probable. And this all has to happen with super-low latency. With the latest NVidia 500W GPUs for cars we might get far better precision than what was possible with Jetson TX2. You are right Deep Learning alone won't make it reliably, as sometimes output goes completely wrong and simple filtering/averaging past few frames might not be sufficient to prevent your self-driving car becoming a self-flying car. But combining all these inputs together, maybe via another DNN, seems like the best way forward.
Re:MIT 6.S094: Deep Learning for Self-Driving Cars
by skgoa on Thursday January 18, 2018 @06:22AM
(#16176279)
Yes, we are using a fusion of different types of sensor input. I have to be careful to not divulge trade secrets, but I would also add that sensory input (deep learning based or otherwise) is only the very first step in a long and complex chain of tools, algorithms etc. The media and academia are focusing almost exclusively on that step, but it's really only a small piece of the puzzle we are working on.
You are also correct on the performance issues. The DNNs you would require for the very complex decisions are currently too large to be used on automotive ECUs. But we also still have big open questions regarding provability of safety when using neural networks that need to be answered before we can use them for core decision making algorithms.
Re:MIT 6.S094: Deep Learning for Self-Driving Cars
by skgoa on Thursday January 18, 2018 @06:28AM
(#16176298)
TBH it seems more like academia is jumping onto the huge wave of hype that has been rolling these last few years. Lately we (a company working on automated driving amongst other things) have seen a lot of interest in internships, thesis topics etc. from students, but only as long as it includes deep learning. It's no surprise to see academia cater to that demand.
Driving on deep learning alone will work great, most of the time. Some of the time, it will do something completely bogus. Like Tesla's self-crashing cars, which use Mobileye's system to not-recognize obstacles which block part of a lane.
Sometimes, the system will sense an obstacle and won't be able to identify it. That has to be handled safely.
[1] https://www.ted.com/talks/chris_urmson_how_a_driverless_car_...