Levels of autonomy for driverless cars

As driverless car technology progresses, there is considerable misunderstanding about the different flavors of autonomous vehicle technology. When car makers talk about driver assistance systems, they have to address some very different issues than when researchers look towards fully autonomous systems which are capable of completely independent (possibly even off-road) operation. We propose to distinguish the following three different levels of autonomous operation (Assistive, Managed, Independent) and provide an explanation of some of the unique issues that need to be addressed for each of these levels.

Levels of autonomy Characteristics Examples Unique problems
Assistive (lowest) Driver assistance system can perform certain driving tasks autonomously, human driver is always needed. Intelligent cruise control keeps lane, manages speed and brake on highways, intelligent parking, …
Mercedes Distronic Plus, Lexus Advanced Pre-Collision System, Volvo Pedestrian Detection
Switching between human driver and driver assistance system. Understanding driver’s intentions; deciding when to act autonomously without driver consent (e.g. pre-crash braking); deciding when situation is too complex for driver assistance system
Managed (medium) Car drives fully autonomously but relies on regularly updated external knowledge (and possibly services) provided remotely Google driverless car; compares environment to continually updated 3D map from Google servers; car can not miss known stop lights because of the map; data exchange is two-way: driving updates the map; Map may include predefined routes for areas that are difficult to navigate. Operation may be limited to mapped area. Building the initial map, maintaining it, interfacing with third-party data sources to keep map up to date.
Propagating changes to vehicles.
Coping with perceived changes to the environment which are not yet in the shared map (e.g. new construction zone: update map; missing stop light: may need additional (human?) verification to avoid sensing errors)
Independent (high) Car operates fully autonomously and matches human driving ability even in unknown terrain without external communication. Mostly research prototypes, e.g MuCar3. Usage scenarios that may require independence: military, off-road, emergency response. Requires a very high level of contextual knowledge and reasoning.
What is the minimum level of prior map knowledge required for safe operation in normal traffic in this mode?
© 2013 Hars, A.: driverless-future.com

Radar on a chip reduces sensor costs

European research project SUCCESS has developed a low-cost radar system which could help reduce the costs for some of the autonomous car sensors. The tiny chip contains all required high frequency processing components and even includes the antenna! The radar operates in the 122GHz band and has a wavelength of 2.5mm. It can measure distances and velocity of moving objects within a range of up to 5 meters with high precision. The sensor was developed by a consortium of European companies including Bosch, Silicon Radar and the Karlsruhe Institute of Technology. At an estimated price of less than 2 U$ the chip has a wide field of potential applications and may be especially attractive for smaller autonomous devices where it could replace currently used ultrasound sensors.

Perfecting driverless cars on the race track

We just came across a talk by Chris Gerdes at the recent TEDxStanford conference. He is busy developing autonomous race cars. Driving these cars autonomously at their limit at high speed, difficult tracks and on slippery surfaces greatly helps improve the algorithms and could also be important to increase acceptance of this technology.
The team has is also carefully analyzing the behavior of professional race drivers to learn about optimal car handling in critical situations. They have gone so far as to equip race drivers with sensors to continuously measure their cognitive load while on the race track. Gerdes hopes to replicate some of the maneuvers that race drivers handle almost instinctively and with very little cognitive load.

Stanford course on the Future of the Automobile

Intelligent vehicles were the main topic of a 1-unit Stanford course from April to June. The course included guest lectures from Volkswagen Research (pdf), Hyundai (pdf), Volvo (pdf) and Bosch (pdf) and addressed technical, legal and some societal aspects of autonomous vehicle technology.

The course was offered by Sven Beiker and Chris Gerdes, both from Stanford’s Center for Automotive Research (CARS). Some of the key insights provided in the class:

  • Carsharing would benefit a lot – autonomous vehicles could be used by carsharing service providers as soon as 2018
  • Completely autonomous vehicles might be available by 2030 (this somewhat contradicts expected use in carsharing by 2018)
  • Over your lifetime you will spend about 1000 days in a car!
  • Autonomous cars and inter-car communication systems should evolve together; however it is difficult to impose new standards
  • The technology is advancing quickly

It is interesting, however, that the economic and business impacts of driverless technology seem to have been mostly absent from the course presentations (with the exception of a few bullet points in the final session). Issues such as how driverless technology might impact the cost of mobility,  what impacts driverless cars would have on the structure of the car industry, and on new business models and services were not addressed.

Nevertheless this has been an excellent course. I highly recommend taking a look at the  syllabus and the many excellent course presentations in PDF format.



Don’t stop at the stoplight: Intersection management for driverless cars

Driverless cars will fundamentally change mobility in more ways than we can imagine today. Researchers from the University of Texas at Austin have taken a hard look at how driverless cars could best negotiate intersections: The classic stoplight would be highly inefficient in a world comprised of only driverless cars. Therefore they have developed algorithms for managing the flow of cars at busy intersections. Cars would signal their arrival at an intersection to an intersection manager and request to pass the intersection. The intersection manager then looks for conflicts with other cars and allocates a time slot for for passing the intersection at a specified speed. This approach is over a 100 times more efficient than the classic stoplight and could greatly reduce congestion, driving times, and petrol consumption in city traffic. A simulation is shown below:

It will certainly take decades until only driverless cars will roam the streets. But intersection management could be implemented long before that time: Once a significant percentage of cars are autonomous, intersection managers could be added to stoplights and issue permissions to those autonomous cars that are at the front of the queue.

Overall this research by Peter Stone and his co-workers shows that driverless car technology holds much potential for improving traffic flow and reducing resource consumption.

More information: Article, Autonomous Intersection Management web site