In the regular debate between Tesla fans and the self-driving community, one of the biggest issues is that fact that almost every developer except Tesla makes use of somewhat detailed maps of the roads in order to improve their driving on them. Tesla says it uses only navigation style maps like those found already in cars and phones today. Those who don’t use maps feel they will be too costly to scale up and maintain over a whole region or the whole world. But a young startup that sells dashcams and has half a million users believes their cameras can build and maintain maps of everywhere at a very low price.
Maps help cars perceive the road around them and the road they can’t yet see, while Tesla prototypes make serious mistakes they probably wouldn’t make with larger maps. Tesla supporters scoff that groups like Cruise and Waymo only offer drivers in the few cities where they’ve created maps and qualified safety. Tesla FSD, while nowhere near being able to drive alone in any city, hopes to do it all on the (probably distant) day when it can.
The autonomous driving industry was born in the big demanding DARPA situations of the 2000s. The first two high-challenge races discouraged maps: you didn’t even get your address until long before the race started. Despite this, the CMU team, which would go on to win second place at the 2005 festival, attempted to map out each of them. Track on course mastery ahead of time so that once the course is revealed, you’re good to go. During the third festival in 2007, all the winning groups traveled extensively along the short route to map it in detail. There would be no surprises, at least not from the sidewalk.
The cards made a big difference. People praised them for being like an extra sensor. You didn’t want to feel the way, you knew everything beforehand. You can just build your map as computationally as you want, you can just check it, you can just ask humans to make sure it’s correct. While other people think that humans drive without a map, we know that we drive better when we already know the road, and in fact use all the symptoms and marks on the ground that tell us what is coming and what it means – same thing. things a robot does with its map. A smart map doesn’t just show the location of a lane marking, curb, or traffic light, it tells what they mean.
Creating a very smart map took a lot of work at first. At Google, which first tried it, mapping was a specialty. In fact, the team of drivers who built the first car had previously worked on the Streetview and GroundTruth mapping projects. At Google, they knew more about maps than almost anyone else in the world. At Google, if someone said, “This would require us to send cars with maps to every single street in the country,” they would simply say, “We did it last month, we can do it again. “
Early mapping approaches were costly. Normally, a fully equipped car with maps travels the street several times and sees it from other angles. The software would process the data and turn it into a map with artificial intelligence systems, similar to, but much better than, the way a car might simply look to discover the map while driving. That map would then be tested through humans who would make sure it all made sense. If there was something the AI wasn’t sure about, humans would repair it. Then a car would get out and drive down the road with a protective driver, and make sure the map was correct. It was then in condition to be used live. Because of human involvement, this came at a cost, much less expensive than other infrastructure costs, such as building or maintaining a road, and was minuscule when spread out among all the other people who would use the map in that direction until it was needed. Renovated.
Even so, it was work to scale this, even if it was economical. That’s when the concept of crowdsourced maps arose. Tons of cars are driving all the roads all the time, and more and more of them are equipped with sensors — cameras at the very least. As technology improved it got more and more possible to build a good map from this crowd data, or to build an initial map with a sensor car, and then keep it up with the crowd data.
While a car without a map has to decipher the road only from what it sees coming, employing only the processors provided in the car, cars en masse travel the road several times, in any direction, in all lanes, day or night. They see everything from all points of view, allowing for much more reliable mapping.
Moving maps, compiled from the knowledge of millions of cars with their chips and cameras, map road features.
Several companies began using this technique. MobilEye took a focus on building a crowd map using very compact uploads from customer cars. MobilEye has cameras in 50 million cars, though it doesn’t have direct access to the chips because that’s under the control of the car OEM that uses their chips. They also have to depend on the car OEM for data uploads, and as such have designed the system to use very minimal bandwidth to upload new data, and very small maps. Even so, MobilEye has demonstrated the ability to map almost all of Europe in a day, and the rest of it over the course of a month. MobilEye also tracks what human drivers really do (which differs from what the road markings may be telling them) at every location on the road. For example, if a tree blocks the view at an intersection so that everybody pulls forward a bit to get a proper look, their map includes that. If everybody slows down or speeds up in certain sections, they know that too.
Other crowdsourcers come with DeepMap, which was an independent company but was acquired through NVIDIA in 2021 (I consulted them before this acquisition). DeepMap began using map vehicles to create initial maps, but was temporarily adapted to standard car information. .
Nexar is a dash camera company. I met Nexar founder Eran Shir when he was one of our academics at Singularity University in California, and brought him into the world of autonomous driving. Nexar started by creating a WiseTephone app to create a Wise dash cam: simply mount a popular phone to your dashboard or windshield, and it can record what it sees and do things like warn of imaginable collisions. Nexar later expanded to sell a dash cam committed to smart features. These millions of dash cameras can now collect and periodically update data on road conditions, at no cost other than the immediate decrease in the cost of mobile data.
More and more automakers have cars with enough computers and sensors (mainly cameras, and rarely ultrasound) to allow their fleets to frequently gather knowledge about the existing shape and condition of roads. Array This ranges from pothole detection (tracking potholes as cars drive over them). to detect free parking spaces – that’s how close those maps are to real time, all for a nearly reduced cost. Tesla is perhaps the company that best offers road maps for its fleet, and although they claim that they do not use very detailed maps in their FSD and Autopilot systems, many report that they do build cards with additional functions. main points in various complex and problematic places.
A map has become essentially a memory. A car that can hope to drive without a map is a car that makes its own map on the fly while it drives. All the car needs do to have a free map is remember what it learned, and send it to the cloud where it can be aggregated with what all the other cars in its network have learned. To deliberately throw away this information is a poor choice. Everybody will use maps with all the detail that can be useful, because it has become very inexpensive to do so. Some companies will prefer to build maps in advance before sending out vehicles with no human on board. If companies have a way to get maps from regular human driven cars, that will be the way they build those initial maps. If they don’t have such a system, they may send employees out to do a few drives on each road. That’s not free but the cost per mile is modest, particularly the cost per eventual rider-mile. After getting such initial maps, the vehicles themselves will do any map update drives as part of their regular task.
When it comes to a robotaxi, which a company deploys in a city and runs, despite its ability to map on the fly, it would be wise to reduce its risks by knowing what is happening on the road before you can approach it. he. enough to map it. The basis will be low-cost participatory mapping.
One Community. Many Voices. Create a free account to share your thoughts.
Our network aims to connect others through open and thoughtful conversations. We need our readers to share their perspectives and exchange ideas and facts in one space.
In order to do so, please follow the posting rules in our site’s Terms of Service. We’ve summarized some of those key rules below. Simply put, keep it civil.
Your message will be rejected if we notice that it appears to contain:
User accounts will be locked if we become aware that users are engaging in:
So how can you be a user?
Thank you for reading our Community Standards. Read the full list of publication regulations discovered in our site’s terms of use.