Stanley was the autonomous vehicle developed by Stanford's racing team, led by Sebastian Thrun, with Mike Montemerlo, Hendrik Dahlkamp, David Stavens and others, that won the 2005 DARPA Grand Challenge, a 132-mile autonomous-driving race across the Mojave Desert from Primm, Nevada to a finish line near the California border. Stanley completed the course in 6 hours 53 minutes, narrowly beating Carnegie Mellon's Sandstorm (7:04) and H1ghlander (which broke down). It was the first time any autonomous vehicle had completed the full course.
Hardware
Stanley was a Volkswagen Touareg R5 (a stock diesel SUV donated by Volkswagen of America's Electronics Research Lab), modified with:
- Five SICK LMS LIDAR units mounted on the roof, scanning the terrain ahead at slightly different angles to build a 3D point cloud of the road surface up to ~25 m ahead.
- A colour stereo camera for longer-range perception (up to ~70 m).
- A 24 GHz radar for very-long-range obstacle detection.
- Differential GPS with inertial measurement, giving sub-metre localisation.
- Six 1.6 GHz Pentium M Blade computers running Linux in the boot, networked over gigabit Ethernet.
- A drive-by-wire kit for steering, throttle and brake actuation.
Software innovations
Stanley's software stack, described in Thrun et al.'s 2006 Journal of Field Robotics paper, Stanley: The Robot that Won the DARPA Grand Challenge, was a clean, modular pipeline of perception, planning and control. Two innovations stood out:
- The drivable-surface classifier. Stanley used the LIDAR's confirmed safe-path labels (regions the vehicle had just driven through, or that were geometrically flat) as training data for an online CNN-style colour classifier on the camera image, extending the perception range from the LIDAR's 25 m to the camera's 70 m. This was one of the earliest large-scale demonstrations of self-supervised perception in a deployed robotic system.
- Probabilistic terrain modelling. Rather than thresholding LIDAR returns, Stanley maintained a probabilistic model of terrain elevation that explicitly accounted for sensor noise and vehicle pitch, dramatically reducing false-positive obstacle detections that had plagued earlier vehicles.
Historical significance
The Grand Challenge series, the 2004 race in which no vehicle completed more than 7 of the 142 miles, the 2005 race won by Stanley, and the 2007 Urban Challenge won by CMU's Boss, is generally credited with founding the modern autonomous-vehicle industry. Before 2005, autonomous driving was a niche academic topic; within five years it was a multi-billion-dollar industry.
Stanley's team substantially populated the founding generation of self-driving-car companies. Sebastian Thrun went on to lead Google's self-driving project, which became Waymo. Several team members founded Cruise (acquired by GM), Aurora, Nuro, Zoox and other AV companies. Mike Montemerlo became one of the founding engineers at Waymo. Stanley itself is on permanent display at the Smithsonian National Museum of American History in Washington, D.C., next to the Apollo lunar module simulator and the original ENIAC components.
Related terms: Stanford AI Lab
Discussed in:
- Chapter 2: Linear Algebra, AI in the 2000s
- Chapter 9: Neural Networks, Robotics and Autonomous Systems