There are a lot of companies currently working to create software for autonomous vehicles. Software developer AImotive is another one of those companies looking to simplify the way that cars see the road. But one of the big differences between AImotive and most other companies is not just the focus on affordable systems for the road—but Almotive’s change from laser-based lidar to an optical-based recognition system.
So far, lidar has been the golden standard when it comes to sensing and mapping the environment for self-driving cars. Lidar operates by using a laser mounted on a vehicle to to measure distance, height, depth, and more by computing the information returned back to it. This information is then turned into usable data, which the car uses to drive.
AImotive is taking a much more natural—and cost effective—approach to mapping out the roadway. The system, which uses between eight and 12 cameras mounted to its test car, works much like the human eye, as the system utilizes pairs of cameras facing each location to sense depth. The cameras will provide a 360-degree look around the vehicle, which aides in capturing lines on the road, objects in the distance, and other important measurable driving-related sights nearby.
All this is meaningless unless this technology could be either better than lidar (where “better” refers to it being more discrete, more cost effective, or more efficient).
As for cost, it’s a long shot to go in either direction. On one hand, Velodyne makes the top of the line HDL-64E sensor, which costs around $75,000. On the other is the more affordable $8,000 VLP-16 puck. There is another piece of hardware squeezed in the middle of things, at a median cost of $29,900, the HDL-32E sensor is more accurate than the puck, and notably costs nearly four times as much.
To make a more affordable option, we would have to judge the accuracy of the AImotive optical-based system versus the laser-based lidar system. Should the visual system’s accuracy be no better than the puck, the pricing would need to be reflective of this. Currently, the AImotive hardware consists of four Nvidia Titan X graphic cards, each with a street value of around $1,000 per unit. The company is reportedly working on some custom Application-Specific Integrated Circuits (ASICs) which would provide cheaper, more efficient processing power at around $100 or less.
Below is another look at the software from earlier this year. One of the developers talks about each piece of functionality to show just how the cameras work in real time.
If a more open solution is available for manufacturers and hobbyists exist, this leaves room in the market for more affordable options. As consumers, it is only natural for us to be hopeful that the product will be delivered as promised; however, until a finished working demo proving it to be as effective as lidar, or a cost-effective report showing a benefit over the lidar puck, it will be difficult to justify the change.