Elon Musk: Lidar an expensive, unnecessary 'crutch'by Justin King
The executive's latest comments suggest Tesla is sticking to its bet on advanced radar mapping, however the company hasn't yet demonstrated such capabilities.
Elon Musk has reaffirmed his confidence that Tesla can deploy an advanced Level 4/5 autonomous system without requiring lidar.
Lidar has become a point of debate in the industry. Tesla appears to be the only company pursuing full autonomy without relying on lidar. Lidar systems with long enough range and high enough resolution to look down a highway several hundred yards cost thousands of dollars per sensor, and some cars have six or more sensors.
Speaking to analysts during Tesla's recent Q4 call, Musk said "it's pretty obvious that the road system is geared toward passive optical," referring to cameras as passive optical and lidar systems (which emit light) as active optical.
"We have to solve passive optical image recognition, extremely well in order to be able to drive in any given environment and the changing environment," he added. "At the point at which you have solved it extremely well, what is the point in having active optical, meaning lidar ... which cannot read signs? ... In my view, it is a crutch that will drive companies to a local maximum that they will find very difficult to get out of."
Tesla is instead taking the "hard path" of developing a neural net that is capable of advanced image recognition, paired with an "increasingly sophisticated" radar.
"You would want to do active photon generation in the radar frequencies of approximately around 4 millimeters because that is occlusion penetrating," he added. "And you can essentially see through snow, rain, dust, fog, anything. So ... I find it quite puzzling that companies would choose to do an active proton system in the wrong wavelength."
Musk in mid-2016 first mentioned the idea of using radar to create a coarse point cloud similar to lidar. His latest comments include a significant asterisk -- "perhaps I am wrong ... but I am quite certain that I am not" -- that suggests engineers are still working out the details and haven't proven a production-viable concept quite yet.
The first fatal Autopilot crash involved a Model S with first-generation sensor hardware that did not recognize a stationary semi trailer across the highway. The trailer was white against a bright Florida sky, a scenario that exemplifies the challenges in using passive optical alone.
The fatal crash raised questions over the role of radar in such systems. The radar technology that currently enables most automatic braking systems must filter out certain obstacles that would create false engagement, such as overpasses and other stationary objects on the side of the road. It is much easier to avoid false alerts if the system ignores reflections from objects that are not in motion.
Aircraft employ a synthetic aperture radar (SAR) system to scan the ground below and create a 3D map. The motion of the aircraft over the target creates the "synthetic aperture" and delivers higher resolution. Other military technologies, such as phased array radar, seem more suitable to forward-facing scans on a car, but Musk has claimed his vision for lidar-free Level 5 does not require more than off-the-shelf radar systems and cameras.
Musk's comments that "we must solve passive optical image recognition" suggest Tesla and other players, including apparent leader Waymo, are still working to develop an AI hardware/software platform that can deliver on the grand vision.
Tesla fell behind its promise to demonstrate a coast-to-coast autonomous drive before the end of 2017. Musk says the company could have done it, but it would have required a bit of cheating with specialized code that would not have worked on any other route.
"I am pretty excited about how much progress we're making on the neural net front," he added. "It's also one of those things that's kind of exponential where ... it doesn't seem like much progress, it doesn't seem like much progress, and suddenly 'wow.' That's been my observation generally with AI stuff."
The executive is nonetheless hopeful Tesla can do the coast-to-coast drive in three to six months.