Last week, General Motors stepped up its autonomous car effort by augmenting its artificial intelligence unit, Cruise Automation, with greater perception capabilities through the acquisition of LIDAR (Light Imaging, Detection, And Ranging) technology company Strobe. Cruise was purchased with great fanfare last year by GM for a billion dollars. Strobe’s unique value proposition is shrinking its optical arrays to the size of microchip, thereby substantially reducing costs of a traditionally expensive sensor that is critical for autonomous vehicles measuring the distances of objects on the road. Cruise CEO Kyle Vogt wrote last week on Medium that “Strobe’s new chip-scale LIDAR technology will significantly enhance the capabilities of our self-driving cars. But perhaps more importantly, by collapsing the entire sensor down to a single chip, we’ll reduce the cost of each LIDAR on our self-driving cars by 99%.” 

GM is not the first Detroit automaker aiming to reduce the costs of sensors on the road; last year Ford invested $150 million in Velodyne, the leading LIDAR company on the market. Velodyne is best known for its rotation sensor that is often mistaken for a siren on top of the car. In describing the transaction, Raj Nair, Ford’s Executive Vice President, Product Development and Chief Technical Officer, said “From the very beginning of our autonomous vehicle program, we saw LIDAR as a key enabler due to its sensing capabilities and how it complements radar and cameras. Ford has a long-standing relationship with Velodyne and our investment is a clear sign of our commitment to making autonomous vehicles available for consumers around the world.” As the race heats up for competing perception technologies, LIDAR startups is already a crowded field with eight other companies (below) competing to become the standard vision for autonomous driving.

Walking the halls of Columbia University’s engineering school last week, I visited a number of the robotic labs working on the next generation of sensing technology. Dr. Peter Allen, Professor of Computer Science, is the founder of the Columbia Grasp Database, whimsically called GraspIt!, that enables robots to better recognize and pickup everyday objects. GraspIt! provides “an architecture to enable robotic grasp planning via shape completion.” The open source GraspIt! database has over 440,000 3D representations of household articles from varying viewpoints, which makes up its “3D convolutional neural network (CNN).” According to the Lab’s IEEE paper published earlier this year, the CNN is able to serve up “a 2.5D pointcloud” capture of “a single point of view” of each item, which then “fills in the occluded regions of the scene, allowing grasps to be planned and executed on the completed object” (see diagram below). As Dr. Allen demonstrated last week, the CNN is able to perform as successfully in live scenarios with a robots “seeing” an object for the first time, as it does in computer simulations.