In an interview, Anshuman Saxena, Vice President and Head of ADAS/Autonomous Driving Products at Qualcomm, provided us with further background information and a look ahead to the future. Snapdragon Ride has a history: According to Saxena, a large global team is behind it. He himself has been with Qualcomm for eleven years, and sometime between 2014 and 2015 it was decided to focus all energy on safety products (and safety-oriented products), which was also the kickoff for the Snapdragon Ride Pilot.
The Snapdragon platform is intended to make driving safer, but what good is that if all users always switch off the ADAS systems first?
Saxena: From a customer’s perspective I can fully understand this: If ADAS systems do not operate robustly and continuously trigger false alarms, people simply switch them off. But the good news is that the new systems are much better! And that is also necessary, because lane-keeping assistants or emergency braking assistants that can save lives are often switched off. The fact, however, is that in the future there must be a justifiable reason for false alerts, so that the user does not lose trust in the system.
How can these be avoided or at least reduced?
Saxena: We now have Level 2 for many vehicles, and many “Plus”, “Plusplus” or “Plusplusplus” extension levels will follow. These are not Eyes-off systems yet. It is important, however, that ADAS systems become more “human,” then people will trust them and they will be acceptable. A good example is the BMW iX3, which was also tested for how users use it in everyday life, including looking in the mirror to confirm a lane change, which is now absolutely “harmonious.” This was tested over millions of kilometers.
How will the systems develop in the future?
Saxena: Intelligence in the car is very important. In the future the car will be able to interact with the user and explain why it is doing something – also because it learns from the user.
Do you have a concrete example for that?
Saxena: Take a tight work zone on the highway, where you always drive a bit to the left in the left lane to overtake the slower trucks with their wide trailers. If you can teach the system to understand why you always drive offset to the left and that you don’t have to stay behind
the trucks where forward visibility can be impaired, it can be wonderfully relaxing in hands-off mode. That’s why in our collaboration with BMW we work with fleet data. I know this from the stretch from San Diego to Los Angeles, where it can be wonderfully relaxing to drive hands-free.
There are now many videos from China with Level-3 vehicles. Is there increasing pressure from this region to take the next step here?
Saxena: We see worldwide strong demand for Level-2 and Level-2 Plus systems. Regulations are also increasing step by step. But this is not specifically Chinese: The BMW solution we are about to present must work worldwide, and as the market develops so do the regulations. But the “pressure” to advance the technology continues worldwide. We are developing similar platforms and software, but the data flywheel is basically the same worldwide. (An AI data flywheel is a self-improving loop in which data from fleet-wide interactions or processes are used to continually refine AI models, generating better results and more valuable data for further improvements. Editor’s note). For legal reasons we sometimes cannot share the data internationally, but the development approach is the same worldwide.
How future-proof is the Snapdragon Ride Pilot? And when will a completely new generation be required on the hardware or software side?
Saxena: That is a good question. The technology must always be future-proof! But not everyone needs all data immediately; they are provided as needed. This is a development, and the automakers have their own plans for it. In practice this evolves constantly, based on the data collected and the scenarios the vehicle encounters. This is fed into the big flywheel, then you train it and reach the next step or the next version. One thing is clear: the more data we have, the faster and more precisely we can do this, and it may well be that in three or five years we again hit the limits of the hardware, then we will have to change it. The more data we have, the faster the development will progress! The first new implementations are already appearing. It is an ongoing process.
So far, complex solutions have been reserved for expensive vehicles; what about cheaper models?
Saxena: We can adapt this, because the overall system is flexible and scalable accordingly. Fundamentally, our applications come from the same software development
and can be adapted to customer requirements. Some things, however, are stringent requirements: For good data you need a good and complete sensor set! With a single camera you cannot achieve what other models with five, six or nine cameras accomplish. The sensor technology determines the operating domain (ODD), the vehicle control applications are then adjusted accordingly.
And can all upgrades be performed over the air?
Saxena: The system is designed so that the software can be updated at any time. It is important to design the sensor and software architecture in such a way that it covers more than the market demands today. The key is not to always touch the underlying base software, but only the applications above it.
What about cybersecurity? Jaguar Land Rover recently faced severe cyberattacks?
Saxena: A very good question! Ideally, you try to exchange as little data as possible within the individual layers through isolation and encryption. At Qualcomm we have several cybersecurity levels as solutions: hardware, Driveware, software. Then there is a further subdivision and several partitions in the system. That means there are various “barriers” built into the system. Fundamentally, of course, cars that are not connected are safer.
That’s our starting point – but wouldn’t that be a step back to the Stone Age?
Saxena: It is wrong to stop the technology; instead it must be developed more securely. All safety checks must be implemented from the start.
How big do you think the step from Level 2 with its “Plus” versions is to Level 3?
Saxena: It is very substantial, because we are talking here about the “eyes-off mode” on highways and not only about low speeds behind lead vehicles. Therefore Level 3, by design, is much more expensive, because there must always be redundancies: practically everything requires redundancies, and that costs money, which is why you cannot integrate it into every car. With Level 3 the question always arises: does it really bring me something and does it stay robust? Driving only very close behind someone on the highway at low speeds does not really offer a great added value. But the development is advancing toward higher speeds and pursuing your map goal, and I am sure that in the future we will see many more L3 systems, even if they will be at much smaller volumes than L2+ platforms. The question
always arises: Do I need to develop something completely new, or can I reach a Level-3 system by continuing to evolve L2+ step by step? And: Is that suitable for the mass market?
What concrete use cases do you see?
Saxena: There are applications where you can benefit especially in the commercial sector: You can autonomously operate in pits or mines, on fields or construction sites, and for that you need specialized capabilities, but that is something completely different from cars, which we have discussed so far. The systems will also be very useful for parking and will soon be available faster and in greater numbers than expected. However, it is always important to consider the overall system.
In what way?
Saxena: If I simply deploy huge computers or computing power everywhere, that does not help. We must always consider the hardware-software coordination; the platform must be adaptable to vehicle classes. We always need adaptive layers to achieve this. For example, if in cheaper vehicles you only have one or two cameras, I must have a parameter in the Ride-Pilot system to adjust it accordingly. The systems also need to be trained, which takes us back to the beginning: We must teach the vehicles exactly how typical behavior patterns in difficult use cases play out: How to turn left in difficult, unprotected curves, how to avoid hard-to-classify obstacles, how to detect whether vehicles in the second row are parked or just paused? All cars collect this data – and this then flows into the AI planner, which helps us achieve more natural driving behavior. We have millions of kilometers from current cars, and the commercial truck and bus fleets have been collecting far more data for a long time, and previously millions of kilometers had to be collected to even implement the systems.
Will there ever be an end?
Saxena: Yes, through deterministic safety measures and repeatability. If a million vehicles at a critical point have reacted almost identically 999,999 times, you can assume that this part is “safe,” and the assessment becomes easier.
What do you see as the biggest challenge for the future?
Saxena: The biggest challenge will be to adapt and scale the system for all different automakers and to deploy it individually. That is our task and our value proposition.
the interview was conducted by Gregor Soller, Editor-in-Chief of VISION Mobility
