Self driving is coming Tesla’s Approach to Autonomous Driving Could Be the Only Scalable Solution
By Tasha Kenney | @TashaARK
This week, Tesla’s AI expert, Andrej Karpathy, spoke at a virtual computer vision event on the Scalability of Autonomous Driving. Karpathy suggested that Tesla’s camera-based approach to autonomous driving might be the only way to build an autonomous system that scales globally.
Karpathy contrasted Tesla’s approach to that of Waymo which relies on LiDAR to map every road down to centimeters of accuracy. Waymo compares the maps to real-time road conditions before making driving decisions. Tesla does not rely on highly accurate maps, preferring cameras, radar, and real-world driving data to direct its autonomous system. With a significant data advantage over its competitors, Tesla can call on data from billions of real-world driving miles while Waymo, Cruise Automation, and others have spent a decade accumulating only 20 million miles.
Tesla’s approach to autonomous development seems to have been the more difficult path, feasible now only because of its data advantage. Waymo and others, meanwhile, will have to map every road before launching their autonomous systems nationally or globally, quite a daunting task! Ford’s Active Drive Assist technology announced this week, for example, will enable autonomous driving on only 100,000 highway miles.
Karpathy also explained that Tesla’s full self-driving solution will not work perfectly in every driving scenario, an important admission because half of the solution is understanding the problem. If Autopilot is unable to drive autonomously through an intersection, for example, it will choose an alternate route that the system understands. Alternatively, and according to ARK’s forecasts, cars will navigate autonomously for the majority of time but will be able to call on “remote drivers” to guide them in rare scenarios. In other words, autonomy during the next 5-10 years will not be level 5, perfectly working at all times.
Even with rare human help, ARK’s research suggests autonomous taxis could cost consumers only 25 cents per mile at scale, roughly a third the 75 cents to drive a personal car or roughly a tenth the $2.50 to take an Uber. In other words, even if not perfect, Tesla’s autonomous taxi network is likely to be more cost-effective than human-driven transportation alternatives and could become a highly attractive business opportunity.