Scale has launched a new mapping tool/service offering for not just self-driving cars, but also for delivery companies. They will now compete with Nvidia, which recently acquired DeepMap, a leading company in that space. We earlier reported on rumors of this service.
Scale Mapping will not deliver maps the way companies like Here and TomTom do, but rather assist clients in creating their own maps with a combination of AI tools and human tagging. Scale announced that Nuro would be a client for this service.
Scale is a leading company in tagging and annotation of sensor data to help train machine learning systems such as neural networks. They employ large numbers of taggers, typically in countries with low costs of labor, who use human intelligence to correctly label targets in things like camera images. Once humans have identified what pixels are cars or pedestrians or lane markers, computers can be trained to learn how to do that. Typically automated tagging starts the process, and humans improve those AI guesses to get good quality training data.
In mapping, however, in order to reduce the cost of making and maintaining maps, most efforts have been to take the humans out of the loop as much as possible. They still play an important role in improving the quality of the final result to a level beyond what computers can do, but an ideal system balances the cost of humans with the quality they can provide. Scale’s main business is human labeling and their offering appears to focus on that.
There are many applications for highly detailed maps, including of course self-driving vehicles. They were pioneered at the DARPA Urban Challenge and later by Google Chauffeur, now known as Waymo. They are used by almost all self-driving teams, with the famous exception of Tesla and a few others. These maps are also useful even for ADAS “autopilot” systems, making them more accurate, though at a cost. It is likely that had Tesla had such maps, they would have prevented some of the fatal accidents which have happened to inattentive drivers using Tesla Autopilot. Maps can also be very useful in creating the simulations that all teams use to test their software.
Scale is also approaching a new marketplace by mapping things like delivery locations and pick-up/drop-off locations. Good maps of these are essential for delivery robots (like Nuro) and robotaxis, but they are also handy for human-driven delivery vehicles and taxis, who want to make sure they take the right path to reach these spots. After all, even if you specify a destination address, the actual stopping spot can be other than right in the center of the property, and you want to know where it is before you get there. The use of human labelers here makes sense, as sometimes even human drivers take time figuring out just what the right spots are.
We also reported yesterday on Pony.AI’s unmanned taxi tests in Fremont, California. They also released a video of such tests in China, in much more complex driving situations.
Waymo released a tweet confirming that their brightly colored vehicles in San Francisco are indeed a Gay Pride declaration, as reported here earlier this week.