A new breed of self-driven cars have abandoned the robotics approach for reinforcement learning with surprising results
Ever since the first experimental autonomous vehicle (AV) capable of driving on and off roads was developed in 1984 by DARPA (Defense Advanced Research Projects Agency) – the most prominent research organization of the US Department of Defense – investors have sunk more than $100 billion into AV research. That’s a third of NASA’s spending for the manned moon mission! Yet, self-driving cars still remain in the research phase. Real-life results are too limited when judged against the quantum of funding and hours of R&D.
Now, a host of AV start-ups are challenging the traditional research approach to driverless vehicles. Armed with pathbreaking advances in AI and reinforcement learning, they are charting a new course in which a new breed of self-driving cars could soon overtake the performance shown by traditional technology.
In 2004, DARPA organized a 240 km Grand Challenge in the Mojave Desert. It was the first long-distance competition for driverless cars in the world. None of the participating vehicles could finish the route. The farthest distance was covered by the entry from Carnegie Mellon University – a paltry 11.78 km!
The 2007 version of the competition was named DARPA Urban Challenge. Organized on the grounds of an abandoned US Air Force Base in California, the course involved a 96 km mock-up set of a small town. The competing robotic vehicles had to finish the urban street set-up within 6 hours obeying all traffic regulations while negotiating with other traffic and obstacles at the same time. Six teams of researchers managed to get their robotic vehicles successfully to navigate the full course – the first three within four-and-a-half hours. That was the Eureka Moment for the AV industry!
Back then, self-driving cars were all about robotics. This was the approach taken by the six winning teams – and ever since all focus of the AV industry has been concentrated on it. The robotics approach treats perception, decision-making, and vehicle control as different problems, with different modules for each. While this offers the advantage of modular expansion, this makes the overall system too complex for development, integration, and maintenance.
Errors in one module could hinder others – leading to component incompatibility. This is a standard approach for automated robots, but real-life driving really does not work that way. Unlike automata, human drivers employ all of their faculties together and that is where intelligence beats robots.
Riding on the success of the 2007 DARPA Urban Challenge, the AV industry saw the emergence of the two most successful names till date – Cruise and Waymo. Both used the robotics approach to the hilt. Headquartered in San Francisco, Cruise LLC is an American self-driving car company. General Motors acquired the company for an undisclosed amount in 2016. Cruise is one of the most advanced self-driving manufacturers. Cruise has built a vast virtual factory to support its software, with hundreds of engineers working on different parts of the pipeline.
Waymo LLC, the other pioneering American autonomous vehicle company, is owned by Alphabet Inc, the Google parent company. Waymo also develops driving technology for use in other vehicles, including delivery vans and tractor-trailers. Till now, Cruise and Waymo are the only AV companies that have come close to successful real-life implementation. While Cruise operates a live robo-taxi service in San Francisco, Waymo’s commercial self-driving taxi service called “Waymo One” operates in the greater Phoenix area of Arizona.
The interesting thing to note is that both Cruise and Waymo vehicles operate within the limited confines of specific cities. And therein lies the shortcoming of the traditional robotics-based AV technology.
Meticulous mapping is a big hurdle
Traditional self-driving cars use high-definition 3D maps to plot the area within which they are supposed to operate. These are created by combining satellite imagery with road data collected by car-mounted cameras and sensors. Such maps refine raw sensor data that the vehicle receives on the go by providing location-specific extra information– for example, lane boundaries, traffic lights, or street curbs.
This process has been undertaken by all major AV manufacturers – resulting in millions of road lengths being mapped throughout America and Europe. There are specialist firms that create and maintain such HD maps which most AV companies subscribe to. An industry leader like Cruise develops its own map – which is definitely an edge over the competition.
But meticulous mapping is a big hurdle. Before a Cruise AV is introduced in a new city, the company must first map all the street details centimeter-by-centimeter. They have to virtually recreate the entire city – including street layouts and every anticipable driving condition – through these maps. This means vehicles that depend on this technology will never be able to navigate an unmapped region as competently. And even for a mapped territory, road layouts can change, requiring maps to be updated endlessly.
A different approach
These are the reasons why a new breed of driverless carmakers have abandoned both the module-based robotics approach, as well as the cumbersome map-making and map-feeding routine. Instead, they are experimenting with neural networks – with surprising results!
But more of that in the next episode.
(To be continued)
Know more about the syllabus and placement record of our Top Ranked Data Science Course in Kolkata, Data Science course in Bangalore, Data Science course in Hyderabad, and Data Science course in Chennai.