Meet ALVINN, the Self-Driving Car From 1989
ALVINN was a neural network powered autonomous vehicle
As the Udacity community continues to make amazing progress on our open source self-driving car, I couldn’t help but Tweet an exciting video a team of Udacity students shared with me. The video demos a student-created deep learning model, which aims to steer a car autonomously using only a camera.
How does it work? After training the model on tons of human-generated driving data, you feed the model a camera image frame (in this case, a series of frames from a trip from Mountain View to Half Moon Bay, a route it’s never seen before) and the model outputs a value it believes would steer the car correctly on its route. You can read more about this very fun problem here. This deep learning focussed approach has gained a lot of traction with those working on self-driving cars as of late. A huge benefit is that you don’t define rules (lane markings, etc.), meaning that given enough (good) training data, the car aims to handle environments it’s never seen before, just like a human would (using past experience).
However, this approach is not necessarily new, which I was reminded of when Dean Pomerleau replied to my Tweet. Dean led a project at CMU called ALVINN (Autonomous Land Vehicle In a Neural Network), a groundbreaking project from the late 80s to build an autonomous vehicle powered by a neural network. I highly recommend checking out the below video, shared by Dean from 1992.
The neural network powering ALVINN was beautifully implemented, but constrained very much so by the hardware. ALVINNs original top speed was 3.5mph, which was limited by the amount of computing power they could fit in the vehicle. In the Udacity Self-Driving Car, we are lucky to have a machine with an NVIDIA Titan X Pascal GPU (>$1000), powering many of our intensive calculations. ALVINN however…
However, after much iteration, ALVINN (read more on NAVLAB) was able to hit 70mph in the early 90s. Watch the below news video from the time for more info, or read this paper for performance, system and hardware details (highly recommended).
I also highly recommend reading this paper on ALVINN, and reading Dean’s thesis he published in 1993. Dean and his team solved many of the problems neural network powered self-driving cars are facing today. Dean also shared a chapter from his thesis on Twitter, about how important it is to generate training data that demonstrates the car recovering from perilous situations.
Dean’s conclusion in the paper is especially prescient.
More extensive testing must be performed before definitive conclusions can be drawn concerning the performance of ALVINN versus other road followers. We are optimistic concerning the eventual contributions neural networks will make to the area of autonomous navigation. But perhaps just as interesting are the possibilities of contributions in the other direction. We hope that exploring autonomous navigation, and in particular some of the extensions outlined in this paper, will have a significant impact on the field of neural networks. We certainly believe it is important to begin researching and evaluating neural networks in real world situations, and we think autonomous navigation is an interesting application for such an approach.
The Udacity Self-Driving Car team hopes we can bring to production, with the help of our amazing students, what ALVINN pioneered in the early 80s. Thank you Dean and team!
