CHANNELS

Aerospace

Automotive

Automated & Autonomous Vehicles

Battery & Electrification Technology

Defense

Drone TV

Electronics

Manufacturing & Prototyping

Materials

Medical

Motion Control

RF & Microwave Technology

Robotics & Automation

Sensors & IoT

Test & Measurement

Aerospace
Automotive
Automated & Autonomous Vehicles
Battery & Electrification Technology
Defense
Drone TV
Electronics
Manufacturing & Prototyping
Materials
Medical
Motion Control
RF & Microwave Technology
Robotics & Automation
Sensors & IoT
Test & Measurement
Aerospace & Defense
Search Results

Remember Me Fondly: Cornell Gives Self-Driving Cars Memories

Unlike humans, cars using artificial neural networks have no memory of the past — despite how many times prior they’ve driven down a particular road. Now, Cornell researchers have produced three recent papers on the ability of autonomous vehicles to use past travels to “learn the way” to familiar destinations. Watch this video to learn more.

“The fundamental question is, can we learn from repeated traversals?” said senior author Kilian Weinberger  , professor of computer science in Cornell Bowers CIS. “For example, a car may mistake a weirdly shaped tree for a pedestrian the first time its laser scanner perceives it from a distance, but once it is close enough, the object category will become clear. So the second time you drive past the very same tree, even in fog or snow, you would hope that the car has now learned to recognize it correctly.”

Topics:
ADAS Automotive Autonomy

More From SAE Media Group

    Transcript

    00:00:03 A lot of groups are working on self-driving cars, and the car is equipped with some sort of sensor that can sense the environment. For the purpose of driving it's not just enough to have the sensing. The car also needs to know, sort of, what are the different participants of traffic like other cars, pedestrians, cyclists and so forth. The way it's typically done is that basically these algorithms are trained and then once they're trained to a satisfactory level, the car is then shipped and the user uses the car. And at this point there's no more learning.

    00:00:37 We just use the algorithm to detect pedestrians, to detect cars, but we are not getting any better at it over time, which is very different from humans, right? The more you drive a car, the better you get at it right? The original idea was from Killian. He asked us one question: "if we can have multiple traversals of the same route over and over again, how can we use such unlabeled data to improve the perception algorithm?" I got this question from Killian and I think it should work in some way because the information's out there and the question is how to use those information.

    00:01:27 I think the most exciting thing is we just started this. I feel like this is really a eureka moment that you can actually utilize this data, which is essentially for free. And that's really the beautiful thing. Like if you are a car company, your customers will be driving around all day collecting data of exactly this format. They're driving around the same routes over and over again. And so you can now use this data to make your cars better and better. One thing that is usually done is the self-driving companies

    00:02:06 are all in California. Which part of the reason is due to the technology, because LIDAR is inherently light. So if you have a lot of extreme weather conditions, you end up with technological failures. But what that also means is all the data that you collect are going to be sunny, clear weather. This is one aspect of our dataset that is pretty unique because we bring in all the stuff, including rain and snow and nighttime conditions.

    00:02:38 So this is something that the community, especially industry, has not really explored. Traffic accidents are one of the leading causes of non-natural deaths in America and around the world. Our research has the potential to reduce that number significantly and make our roads a safer place.