A frequent topic on these pages concerns Silicon Valley and its ongoing attempt to disrupt the car industry. But over in Europe, a former Ferrari F1 engineer has some thoughts of his own on the matter and thinks his company—Cogisen—has a better way of doing things. Christiaan Erik Rijnders spent several years in the early 2000s working on Ferrari's simulator, simulations, and vehicle dynamics, during which time Ferrari utterly dominated the sport.
Through that work, he absorbed a few important lessons on the way successful R&D programs should run, and they're lessons he's been applying more recently to the problem of image processing. In particular, Cogisen has some very cool gaze-tracking algorithms (earning the startup a Horizon 2020 grant from the European Union) that may have an important role to play in autonomous vehicles. We spoke to Rijnders recently to find out more about his time at the Scuderia and to talk about where he thinks most traditional tech startups are going wrong.
"On the inside [of Ferrari], it was everything it was cracked up to be; you really got to see what proper engineering is and proper management is," he told Ars. "You'd have world-class engineers who were all very strong in theoretical fundamentals of what engineering is, who would all share data in a climate where risk taking was encouraged, where nobody was afraid to make decisions and they were made very well, especially on a strategic level by Ross Brawn, on a political level by Jean Todt, and on a technical level by Rory Byrne."
Rijnders stressed the importance of that decision-making process. "Imagine that you might have 10 different development options in front of you. Rory Byrne at a very early stage could point his finger where we should concentrate our efforts, and he was almost always right. This means an incredible efficiency for development," he explained. The alternative approach—setting up 10 different R&D programs to investigate each option—often ends up with each team defending their hard work irrespective of whether or not it's actually the best option for the company.
While at Ferrari, Rijnders worked on designing the simulations and simulators for the team. "Today's simulation engineer opens the simulator package and enters the numbers and runs the program. Back then, you had to create the simulations and simulators, so you really had to go into the fundamentals," he said. Ferrari quickly realized that the traditional method—cramming sensors onto a car and then trying to approximate a model—would no longer suffice.
"You can't do steady state approximations; you have to do something much more realistic. It's really the dynamics and the fact that it's a transient non-linear machine on the limit, and there aren't two fractions of a second where the car is on the limit doing the same thing," he explained. Working on new generations of algorithms that allowed accurate modeling of the car's behavior on track left Rijnders with a realization that there was a similar need coming in image processing.
"At that point with Cogisen, we started working on applying these things to image processing, like detecting pedestrians, detecting where you're looking at without the use of extra hardware. Very difficult non-linear problems. And to develop these things, I applied the same techniques I learned at Ferrari," he said.
It's also where he finds fault with the traditional startup culture. "Sometimes in tech startups, you see people think, 'We're a startup, so we can do things in a hacky culture,' but that's not true at all," he said. "You need to have the proper engineering and management methodology. You have to be brave making decisions in the same way Rory Byrne did. You have to have a high-quality engineering [team] with a lot of practical experience and understand the fundamentals. A lot of those same things we apply in Cogisen."
Interestingly, Rijnders told us that at Ferrari, invention was a repeatable process (something that makes perfect sense when you consider how much development occurs in each of the top teams during the course of an F1 season). And he fears that's not the case with many new tech companies. "To be successful in a technology startup, you have to bring something that's more advanced than the state of the art. More than likely, you're going to have to repeatedly invent something new, so invention needs to be a process. A lot of times now, you see startups that think you can base your whole startup on one brainstorming session and then hack around it. For technology, you have to really do proper engineering methods," he told Ars.
Rijnders was also critical of the rush to autonomous driving, suggesting that lessons already learned in commercial aviation are being overlooked. Referring to the old joke about autopilots and dogs, he pointed out that the reason we still have two humans in the cockpit is to handle the unforeseen events that can't be preprogrammed. In much the same way, it's going to be quite some time, he suggested, before autonomous cars can be expected to handle the Rumsfeldian unknown unknowns.
This is also where Cogisen's gaze-tracking algorithm comes in. "Instead of the driver, we become the nagging mother—'look out on your right, don't drive too fast, overtake that car,' and so on. While airline pilots are needed to make strategic decisions, in the same way, we'll still be needed to make these decisions for autonomous vehicles. Which means an interface of 'how would your nagging mother interface with you while you're driving?' With gestures, voice, and gaze—the way she looks at you. The same interface is needed. You don't really need a steering wheel, but you need somebody to have high-level control of the vehicle," he said.
A version of this idea (although whether it's from Cogisen or not is unknown, since the company is still under NDA with the OEMs it's working with) is due to appear in next year's Audi A8, which the company says will be the first SAE level 3 autonomous vehicle on sale. Competitors like Ford and Volvo have said they plan to skip level 3—which is partial autonomy that can be handed back to the driver on relatively short notice—due to the problems involved in transferring partial control. But Rijnders says that gaze tracking isn't just about knowing when to return control to a human being.
"Gaze tracking won't just be important to give back control to the driver, but [it's] also going to be important in more autonomous vehicles because it will be the nagging mother." He imagined a hypothetical situation where your self-driving vehicle looks like it's about to be hit by another car. "In a dangerous situation, you're never going to say, 'Careful, on your right there's a car coming, and it's going too fast.' You'll shout, 'Look out,' and the car needs to know where you're looking. Even for those steps, I'm convinced that gaze-tracking technology will be necessary."
Rijnders added that gaze-tracking has other important applications, too. "We have to make the difference between infrared technology that simply knows if your eyes are open—drowsiness monitoring—that's one thing; another thing is knowing exactly what direction people are looking at. We do the latter. We really analyze where people are looking."
The CEO compared the situation to Formula 1 of the early 1990s, when teams like Williams were pioneering new ground with active suspension, two-way telemetry, and a host of other technologies that were subsequently banned for being too clever and freeing up some of the drivers' workload in the car. "What we're trying to do with autonomous vehicles is to get people to manage a higher-level view of the car. In the same way, the technology of the early '90s was making the driver get a higher-level view of the car. You still needed a lot of skill to drive those cars fast, because for these cars, there aren't two laps of the year where the car is the same. It's very difficult to get that last percent on the limit, which is where you get speed over your competition," he told us.
reader comments
44