ALVINN
Era 3 · From the expert boom to the statistical turn (1980–2011) › Robotics and perception
- Lineage
- Neurons and networks
- Kind
- Implementation
Who
Carnegie Mellon
What happened
Dean Pomerleau of Carnegie Mellon presents ALVINN (Autonomous Land Vehicle In a Neural Network), a three-layer neural network that learns to follow the road from camera and laser rangefinder images and decides which way to turn the steering wheel.
Why it mattered
A neural network drives a vehicle: it learns the task from examples instead of having it programmed with vision rules.
The full story
The autonomous driving systems of the time, such as those in the DARPA program, analyzed the image with hand-written vision algorithms to find the road and only then decided on a direction. Pomerleau turned the problem around: give a neural network the whole image and ask it directly for the steering angle, with no intermediate steps.
ALVINN had an input layer that took a low-resolution camera image and data from a laser rangefinder, one hidden layer, and an output layer whose units stood for steering directions. It was trained with backpropagation on images of simulated roads and tested on Carnegie Mellon's test vehicle, where it followed the road successfully under certain conditions.
Pomerleau observed that the network's internal representation changed with the training conditions, which suggested a system that adapts its processing to each situation. The paper itself presents the result as evidence that networks trained with backpropagation could handle real perception and control tasks, not just toy problems.
ALVINN is a forerunner of end-to-end driving, in which a single network learns to go from the image to the steering wheel, an idea that deep networks would take up again decades later with far more computing power and far more data.
Videos
- NavLab 1984 - 1994 (opens in a new tab) (CMU Robotics Institute)
Builds on
The Strategic Computing Initiative is cut back (1987)
Documented direct influence
ALVINN was tested on Carnegie Mellon's Navlab, funded by the Strategic Computing Initiative
Leads to
Stanley wins the DARPA Grand Challenge (October 8, 2005)
Inherited, with no proven direct link
The same tradition of autonomous driving built on machine learning, with no proven direct link between the two systems
Threads that pass through here
- Cars that drive themselves (step 3 of 4)
Sources
Original source
Scholarship and context
How to cite
“ALVINN.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/alvinn
This entry was last reviewed on November 13, 2026.