Breakdown Pt. 2 - Maps, Sensors, and Testing

I started these articles to help my network better navigate the Autonomous Driving Industry from hiring and job seeking perspectives. To effectively grow these articles into a discussion of how engineers move within the industry, I'm breaking down the "ecosystem" of technology that is allowing Autonomous Vehicles to move forward quickly. Specifically in this article:

1. HD Mapping Space
2. Camera/sensor companies that enable Lvl. 4 + 5 Autonomy
3. A quick note on the team helping (nearly) everyone test their tech.

HD Mapping

This is one of the most competitive spaces within the ecosystem of autonomous vehicles. There was already an existing market for the tech, with Navteq (now recognized at HERE Technologies) going back to 1985. There was the rise of GPS Mapping devices like TomTom or Garmin, and more recently, smartphone navigation tech like the highly popular Waze (acquired by Google in 2013).

Now, companies are building the most accurate maps yet, using a mix of satellite, camera, and sensor (lidar/radar/ultrasound/you name it) tech. With the combination of these sources, companies are able to get environment perception and localization abilities down to <10cm of accuracy. With Machine Learning they're able to automatically recognize and classify certain objects, from stop signs to potholes. This is crucial to the autonomous driving vehicles.

There are a few different techniques companies can go about this (2 of which broken down nicely by Synced on Medium)

Mapping Solutions Service Providers

The best known of these startup teams would be Deepmap AI. Deepmap is provides mapping solutions as service, outfitting companies with the hardware, software, and data collection management services that allow Autonomous Vehicle companies to transform their fleet data into HD Maps. Their team is built of a strong variety of engineers - expert Sensor Technologists, high-level Machine Learning AI engineers, and hardware experts.

With their solutions providing model, Deepmap has been able to partner with large players in the space, like Ridecell, amongst other Tier 1 (Bosch), OEMS, and large tech players like Nvidia.

They help teams outsource one of the most time and data consuming aspects of autonomy while maintaining a personalized solution that fits perfectly into a client's need. This is not a full-stack solution and some infrastructure is needed by their clientele, but this is how they are able to deliver strong HD Mapping help where they can. From Global News Wire -"Srinivas Reddy Aellala, a leader a Ridecell's Autonomous Driving Group “DeepMap provides best-in-class mapping software features, helping Auro reduce its AV deployment time and effort at complex and remote sites.”

Another aspect to Mapping as a service can be broken down even further, for instance, Point One Navigation providing Localization as a service, helping customers pinpoint exactly where their vehicles are based on GPS and Computer Vision Technolgy.

HERE Technologies, one of the largest and oldest Mapping companies (Navteq->Nokia-> HERE) provides a more full-stack HD mapping solution for their OEM partners like Audi and BMW, providing HD maps for the navigation system. They've partnered with Israeli based Mobil Eye for stronger computer vision based map creation, and Nvidia to increase the use of deep learning. Their base for maps, however, comes from crowdsourced data.

Crowdsourcing Maps

This is a larger part of the HD Mapping space, as many teams have realized the best way to quickly collect data to build the maps is from, well, everyone out there driving! Different teams have taken different approaches. This is one of the cheapest methods of HD Map creation, however, there is a high threshold for the data and requires a lot of post-processing.

Lvl5 - Encourages people to use Payver, an app that makes your smartphone into a dashboard cam, recording your trips allowing them to collect data from your drives. They used to actually pay Uber and Lyft Drivers to use it, but looking at the Payver website, it seems now they are encouraged to use it to protect against liability, keep a trip diary, and well, collect coins.

Anyways, they use the video and data collected from the smartphone to create maps that self-correct and update. Lvl5 Was actually sold just last week, with more details coming out in the coming months.

Civil Maps - Crowdsourced Edge Mapping. One of the problems with crowdsourced mapping data is the heavy post-processing load that occurs once the data is collected. Civil Maps is taking care of this aspect through edge mapping, processing the data closer to where it is captured, in the vehicle, prior to sending to the cloud infrastructure. I recommend checking out this article to really understand it better.

Mapper - Provides maps to AV companies needs from crowdsourced data. They map the area, customize it to the clients "schemas and sensor requirements" and deliver it through simple web API's.

Having accurate and updated maps is one of the most important steps to level 4/5 autonomy. Some teams, like Gatik AI, even specialize in "Geo-Fenced" areas, determining specific routes to operate within based on their client's need. This way they can map a smaller area and deliver autonomy more effectively and at a faster rate than teams who may have to operate within an entire city, state, or country.

Camera and Sensor Companies

One of the most expensive aspects of Autonomous Driving is the system of cameras and sensors each vehicle should have to safely perceive the world around them and navigate autonomously.

Here are the majority listed out:

• Stereoscopic Camera Systems - Not just cameras, but the ability to see 360 degrees around you and stitch those images for a clear view.
• Lidar - Lasers (often rapidly spun 360 degrees) measuring the distance from the time it takes to bounce back to the sensor. Creating a 3D "point cloud" of the surrounding environment. (See photo in the section below)
• Radar and Sonar
• IMUs

Every company has a different opinion on which sensors/cameras are necessary for Lvl 4/5 Autonomy. Tesla's CEO Elon Musk has a pretty strong opinion on his vehicle's choice of sensor. "Lidar is lame", saying it is expensive and pointless. Tesla's autopilot opts for a system of camera, sonar, and radar sensors. But Tesla has always been a bit bolder in their promises to the industry. A lot of other leaders believe Lidar is an absolute necessity for truly safe autonomous navigation, as no other options are as accurate as Lidar. Lidars are extremely expensive though, high-end Lidars running for $75,000+. Ouster says their lidars are only $12,000 per sensor. Still pretty expensive, especially considering some vehicles will need multiple lidars attached.

In turn, the demand for cost-efficient and highly accurate cameras and sensors has created a huge playing field for development.

Lidar

Below are 4 that I see having the largest impact on the SF market for Autonomous Vehicles.

Velodyne - Creator of the first modern Lidar sensors. Started Lidar research in 2005. Some of the most expensive lidars available but also the most powerful on the market.

Ouster - Winner of 2019 CES Innovation award, they are cutting down the price of Lidar while maintaining a high level of accuracy and range. “A Corvette is 90% as fast as a Ferrari and it’s 10% of the cost,” said Angus Pacala, Ouster chief executive and co-founder.

Luminar - Some of the farthest ranging lidar sensors, Luminar is a major competitor to Velodyne. Scored partnerships with Toyota, Volkswagen, and Volvo.

AEye - Very similar to Luminar, they boast high range lidar sensors capable of reaching distances of 1000M.

Here is a breakdown of the 10 leading teams.

Cameras and Computer Vision

Since Lidar is not only expensive and high-powered but also cannot recognize color and fonts that help in classification of stoplights, and road signs. Smart cameras are seen as a valuable asset to lidar's long range and 3D modeling abilities. Cameras also need to be able to see in low light conditions with high accuracy. Here are a few companies that are helping innovate automotive cameras and computer vision algorithms that improve the capabilities of photos taken.

Light - Integrates images from 16 different lenses to create 3D maps that compete with Lidar.

Ambarella - Creating deep learning and optimization software that companies can use to improve existing camera system abilities. They want to help vehicles see under difficult conditions, like dark nights or snowy days.

Deepscale - Using Fullstack Deep Learning solutions, a large part is building algorithms to improve the classification and identification abilities of cameras and other AI applications.

This article hyperlinked in the photo below has a great detailed breakdown of what I just ran over from a higher, more world market-based viewpoint.

R&D Testing

After my last article, I had the pleasure of speaking to someone about AutonomouStuff - a company I had seen around my network for some time but had not been able to get a grasp on what they were doing. AutonomouStuff is in somewhat of a league of their own, working with several of the biggest names in autonomy (like Aurora Innovation, Baidu, Nvidia) while enabling the smaller teams to start gaining valuable insights from their own tech.

AutonomoStuff's wide array of products and solutions help companies gain the architecture and mechanical vehicle systems to start testing out their motion planning and control algorithms, perception algorithms and start building their own maps, amongst many other critical points. They work with teams to deliver test vehicles that clients can outfit with their AV system. Depending on where your team wants to specialize, they can help cover the bases of the rest of the stack to help you develop your specialty - whether that be a specific sensor or a full-stack autonomous driving solution.

"It takes a village"

When the first vehicle reaches fully Level 4 autonomy, it won't be accredited to a single company. Autonomous driving can't be possible without this wide ecosystem of niche-focused tech companies who partner up with the larger teams building the full-stack solutions or enablers that allow quick progress to be made. This is why we see a lot of engineering movement throughout the space as well, as Autonomous Driving offers a lot of different ways for engineers to develop their skillset.

Hope everyone enjoyed the second part to the series. Will be getting going on the next one soon so they can come a bit more consistently. What would you like to learn more about in articles to come?