The rise of autonomous vehicles, calls for improving the systems regularly. Every mile that an autonomous vehicle covers delivers valuable data in their development. But how do we know if the system is ready to be set out in the field?
Puneet Gupta, Chief Technology Officer, Brillio, talks about the challenges involved in autonomous vehicle development with Saurabh Durgapal of EFY.
Q. What are the new trends in automotive testing?
A. In the early stages of system development while building a core system, Its very important to invest in building emulators. This helps in managing testing in the early stages and reduce the cost at later stages. The cost of corrections increases significantly, if you go for testing after the system has been developed. Use of AI in creating user modelling and auto creation of new test scenarios is a high potential area
Q. How does AI come into the picture in testing?
A. While testing any autonomous system manually, the number of test cases are limited by the imagination of the tester in charge, whereas a car can be driven by some body with a much different notion of driving. An AI on the other hand, can model various kinds of drivers based on their driving habits, hence giving a better test case scenario. If we had an efficiency of 60-65 percent with manually testing, adding AI can increase it to about 90 percent.
Q. Could you give us a walkthrough of the system in an autonomous vehicle?
A. There are three major parts in the system involved in autonomous vehicles. The first one is the intra-vehicle sensory network and associated edge processing. This is involved in instantaneous decision making. Second, cloud based analytics and control hub that processes the enormous data being generated and draws insights and behaviour modification decisions. Third is potentially and optionally the external sensor and communication infrastructure that can help enhance the autonomous performance and experiences
Q. How heavily can we rely on middleware?
A. With the right architecture, we can rely heavily on middleware. Everything can be broken down to critical, medium critical and low critical decisions. The critical decisions happen inside the on-board computing device in the vehicle. Decisions that can wait for example, driving behaviour and optimal speed can be sent to the cloud. There is also a possibility of some on-board component failing. The algorithms in this case are designed to take actions depending on the car health.
Q. How do these systems learn to be safer with human life?
A. At the core is learning and risk calculation algorithms. In simple terms, Initially the systems would be overly cautious, which would reduce over time in their tryst to imitate humans. Alternatively, the system can operate in an assisted mode as opposed to autonomous mode to begin with. And once the systems learn to take calculated risks, then system can graduate to an autonomous mode. There is a beauty in which the algorithms are designed to work by carefully weaving in imperfections that result in more intelligent behaviours.
Q. How can modifying algorithms to suit our changing requirements impact the vehicle?
A. Accidents can happen when we try to stretch the boundaries of the algorithms in autonomous systems, trying to make them behave more like human beings. As humans, while driving, we don’t stop at every turn. Instead we take a calculated risk and go ahead. The computing systems are typically built around hard choices, and hence are not tuned to calculated risks. This is where we begin pushing the boundaries of an autonomous system to take some calculated risks.
Q. What are the future expectations in delivering “end-experiences”?
A. Just a smart phone on its own can bring in smart driving experiences today. A range of sensors are already available and more are getting added. So, if the vehicle is involved in an accident, the phone instead of acting dead, can react on the sudden deceleration data and guide you. It can arrange crash pickup or ambulance services as required. It can intimate somebody who can repair the damage. Once the data is accumulated, the focus is going to be heavily on innovative uses of that data.
Q. What is the status of the available analytics tools in regards with testing?
A. There are some good open source and commercial options available today but most of them are in patches and must be clubbed together for extensive usage. So, anybody doing extensive usage must build some core components of their own. That’s the approach we are taking as well. Leverage the good open source available but investing heavily to build the core integrated foundations.
Do you like this interview? You may also like other related articles: click here