Research · 1988

ALVINN: a neural network steers the van

At the NIPS conference of 1988 Dean Pomerleau of Carnegie Mellon described ALVINN, a three-layer back-propagation network that takes a 30 by 32 camera image and an 8 by 32 laser range-finder image (1217 inputs in all) and, through 29 hidden units, outputs one of 45 travel directions. The network was trained on 1200 simulated road images for 40 epochs; on new simulated images it picks the curvature within two units about 90 percent of the time. On the Navlab van it drove the vehicle along a 400 metre path through a wooded part of the campus at half a metre per second.

Why it matters

For the first time the steering of a real vehicle was in the hands of a neural network trained on examples rather than an algorithm a person had written: the processing depends on what the network was trained on, and the weights of the hidden units read as road-edge filters. It is the direct ancestor of driving by camera and learning.

The input is split into two retinas and one feedback unit that says whether the road is lighter than the non-road; the blue band of the image is used as the most contrasting; the desired output is a hill of activation around the curvature that would bring the vehicle to the road centre 7 metres ahead. The simulated images come from the group's own road generator with varied positions, lighting and noise, because collecting real ones under every condition was hard. Navlab is a modified Chevy van with three Sun computers, a Warp processor, a video camera and a laser range finder. Under the same conditions on the same course the ALV group at CMU reached the same accuracy at 1 m/s by running its algorithm on the Warp; ALVINN was then simulated on one on-board Sun, and the author expected speed-ups on the Warp. The proceedings are Advances in Neural Information Processing Systems 1 (NIPS 1988), pages 305-313. What the record does not claim: the month of the conference; kilometres (the paper gives 400 metres); that the network was trained on real images (the paper says simulated ones).

Event record

Event date

1988

Timeline date

Event date

Verification

Sources gathered automatically · September 23, 2026

Lines

Autonomous driving

ID

evt-0650

The proceedings volume is filed as Advances in Neural Information Processing Systems 1 (NIPS 1988); the sources read do not name the month of the conference. The CMU page gives December 1989 in its BibTeX, which is not the conference date and was not used.

Sources

primary ALVINN: An Autonomous Land Vehicle in a Neural Network
NeurIPS Proceedings · Published 1988

Related events

Builds on Navlab: a van drives a road without stopping
The network was built to control Navlab itself, which the paper calls the test-bed of the CMU autonomous land vehicle project, and compares with its traditional algorithms on the same course.
ALVINN: An Autonomous Land Vehicle in a Neural Network
Contrasts VaMoRs drives the Autobahn at 96 km/h
The paper sets itself against systems whose processing stays fixed across situations, naming Dickmanns and Zapp among them: there a road model and recursive estimation, here weights learned from images.
ALVINN: An Autonomous Land Vehicle in a Neural Network
Related Tesla Autopilot
The same idea, a camera and a network trained on examples in place of written rules, became a production autopilot twenty-seven years later.
Related Google reveals its self-driving project
Google's road vehicles grew out of the DARPA winners, and those out of the Navlab and ALVINN line at Carnegie Mellon.

Records that link to this one

Enables Warp: a systolic computer in a van
ALVINN's network, by its paper, was trained with the Warp back-propagation simulator; the paper on that simulator gives 17 million connections per second.
Neural Network Simulation at Warp Speed: How We Got 17 Million Connections Per Second (Dean A. Pomerleau, George L. Gusciora, David S. Touretzky, H. T. Kung), IEEE International Conference on Neural Networks, San Diego, 1988, II-143-150 ALVINN: An Autonomous Land Vehicle in a Neural Network