Autonomous Cars—Are We There Yet?

This is the second of two posts on the CASPA Symposium on Autonomous Driving. The first was CASPA Autonomous Driving 1 . Chia-Lin Simmons Chia-Lin talked on Are We There Yet? Autonomous Cars and Legal Considerations . If Max suffered from the "if all you have is a hammer, everything looks like a nail" syndrome, seeing the cloud in everything, Chia-Lin sees legal issues lurking behind every technical advance. But despite that rather dry introduction, she had some interesting issues to discuss. She pointed out that it took a long time to make state driver licenses work across the whole US, and it is even more difficult in Europe with different countries. Something as seeming simple as streaming media in your infotainment system requires legal stuff to be solved if the music isn't going to stop when you drive from France to Germany. Today almost all laws are about NHTA level 1 and a little level 2. There is a driver assumed. Level 4 has no legal framework. In Texas, it is the driver's responsibility to make sure everyone is properly restrained; in Massachusetts, it is the driver's duty to fill in a form after an accident. What happens when the car is driving a kid to school? Autonomous vehicles can stop on a dime compared to human drivers. Should they have different speed limits? Do we post them? There will be people killed by autonomous vehicles. Even if the accident rate goes from 40,000 to 50, those 50 will all sue the car manufacturers. How do we stop car manufacturers spending their whole time in court and encourage innovation? A model might be what we did with vaccines. About 1 in 100,000 kids inoculated has a reaction. Before we had strict reliability, they were going to court every day, nobody wanted to be in the vaccine business: people die, they die tragically, they are kids, they get big awards. In the end, the US passed a law that all vaccines are taxed, the tax goes into a kitty and payments come out of it. Maybe autonomous cars need to be like that. Insurance is another area where the way we do it may no longer be appropriate. It is tied to the driver (age, sex, experience, track record). It is a $228B business. With very rare accidents, does it go away? Is is legal to discriminate between autonomous and conventional vehicles, given that it will initially be discrimination between rich and poor? Modar Alauoi Modar, the CEO of Eyeris talked on Applications of Deep Learning-Based Emotion Recognition Software in Autonomous Driving . Never mind Big Brother, your car is going to be watching you. Current work is all on semi-autonomy, what the NTHSA calls HAV for highly-automated-vehicles. Driver monitoring is really important on the transition from automatic to manual driving, and vice versa. Deep learning is used to look at faces and recognize emotions. This was hard work since there are no public datasets so they created their own libraries with millions of images from hundreds of people from five different races, four different age groups, and two genders, plus 10 different lighting conditions, 13 different head poses, and a lot of different types of cameras (RGH, greyscale, near IR, VGA up to 5K resolution). The outcome of all this is the seven universal emotions (joy, surprise, sadness, neutral, anger, disgust, fear). Also, detecting yawning, microsleeps, eye-gaze tracking, and more. When they started in 2014, there was lots of skepticism. Will there even be cameras in cars? What to do about privacy? Fast forward 12 months and they were getting awards. Fast forward another year and they were nominated in more categories (ADAS and safety, usage-based insurance). It is now established that driver-monitoring systems will make it into cars and likely eventually be mandated, especially at level 3 of autonomy. They are working with Honda, Toyota, "a US vehicle manufacturer", premium European manufacturers. In answer to a question about what is recorded, Modar said that no images or videos are recorded or stored anywhere, all the driver monitoring is "processed and shredded in real time." Beyond that, it is not settled. For some insurance companies, some go into an insurance exchange portal. GM announced each car will have an embedded dongle (from 2018). Because of dongle data connecting to insurance exchange, some goes over the cloud, and is used by insurance companies. Averis does the same and some analytics go up to cloud. Banks leasing a car want to know if it is well driven. DOT wants to know locations where drivers are stressed. Many industries are interested in this data, not just the vehicle manufacturer. John Eggert Tony Han of Baidu was meant to be the last speaker, but he couldn't make it, so the day wrapped up with John Eggert of Velodyne Lidar talking about...lidar. Well, he called it The Increasing Role of High-Definition 3D LiDAR in Highly Automated Driving . Velodyne started in 1983 making sub-woofers, and then got into robot wars where they came third worldwide. In the DARPA grand challenge, they came third (although they only managed six miles. (For more details on this, see my post Ten Years Ago Self-driving Cars Couldn't Go Ten Miles .) They realized that lidar was the ideal self-driving car sensor and have been commercializing it since. He described it as basically a laser pointer that measures the time of flight. The innovation is how to do this on multiple channels, rotating 360 degrees. This requires lots of lasers aligned collinearly with variable power. They get a huge dynamic range by varying the power depending on how reflective things are. This lets them build a model of the environment almost instantaneously with 400,000 laser shots in the blink of an eye. By the DARPA urban challenge in 2007, five out of six finishers relied on 3D Lidar. By then, Velodyne was no longer driving their own vehicles but supplying to everyone (including Stanford and CMU who came first and second). Lidar is great for road debris (or "unmapped obstacles" as apparently they are known). They need to see a 20cm object 100m down the road. California is the only place in the world that makes autonomous vehicles supply data and makes it public. For example, how far does a car need to go before the driver has to intervene? For Google, it was 2000km, some others only 20km. By 2016, Google was up to 10,000km. They need to take account of the "track complexity" since there is a big difference between freeway and urban environments. Cruze is only three years old (now part of GM) and is surpassing established OEMs due to lidar. There has been a big change in thinking from "don't need lidar" to the vast majority using it. Only one thinks they can do without (I think that is Tesla, which uses only cameras, although Tony didn't say who). It is used by all the non-traditional upstarts such as Baidu, Navya, Zoox, Nutonomy, Easymile, and many more who are confidential. For level 4, Lidar needs 200m, high resolution, but also small size and low cost. The key constraints are emitter power (heat, eye safety), detector sensitivity (can’t have false positives or disengaging all the time), lens manufacturability (small, short focal length, small field of view). Lidar sends billions of photons and collects just a few. The result is that the sensor fits in a hand, not like those big ones on top of Google's early vehicles. In fact Tony was proud of the progress that has been made in 10 years they have been working on lidar: Size 1m to 100mm Weight 30kg to 803g Cost $85,000 to $500 (mass production) Automotive durability The biggest limitation is fundamental, it depends on the line of sight. The most advanced use are dump trucks in Australia. They have been fully autonomous for years and with a duty cycle nearly 24 hours a day. In the Q&A, Tony was asked how it will work when you have a complex junction with 40 cars all using lidar. He said that is a challenge for the future, today you never see more than five or six, so it is okay for now. In the future, they will require better filtering and they are working on the issue, but it is still confidential. Next Symposium The next CASPA Symposium will be July 8. For now, the topic is secret. For the big reveal, keep an eye on the CASPA website .