Opportunity & Solution Summary

Beyond the enormous societal benefit of reducing traffic collisions, a connected fleet of autonomous vehicles allows for more predictable, efficient traffic flow; improved mobility and productivity among travelers; and–eventually–a business model shift from outright vehicle ownership to ‘transportation-as-a-service’.

Looking ahead, the National Highway Traffic Safety Administration (NHTSA) created a five-level classification system of autonomous capabilities to measure progress and innovation:

In October 2015, Tesla Motors pushed software version 7.0 to its Model S customers, which included Tesla Autopilot, the most advanced publicly-available autonomous driving software.

While many companies have developed autonomous capabilities (particularly Google, who, as the first-mover, logged 1 million fully-autonomous miles before Tesla launched Autopilot), Tesla’s software has uniquely iterated and addressed the changing needs of the user to become the superior solution.  Interestingly, 20+ automakers have more autonomous driving patents than Tesla (mostly surrounding anti-collision and braking control mechanisms), but Tesla has been the first automaker to provide substantial Level 3 features in the marketplace.

This has enabled Tesla to leverage its thousands of drivers to quickly improve its algorithms via ensemble training.  By pushing these solutions to the market, Tesla has logged 50-fold more autonomous miles (supplemented by user feedback) than Google to boost algorithm performance.  In the short run, this means improving vehicle efficiency and customers safety.  In the longer run, this means reaching full self-driving automation (“Level 4”).  The software’s continuous learning technology enables the autonomous cars to update as new processes are observed from the user.

NVIDIA and Tesla together have fed millions of miles worth of driving data and videos to train the computer about driving.  Tesla leverages NVIDIA’s DRIVE PX 2 platform to run an internally-developed neural net for vision, sonar, and radar processing.  DRIVE PX 2 works in combination with version 6.0 of its deep-learning CUDA® Deep Neural Network library (cuDNN) and Tesla’s P100 GPU to detect and classify objects 10x faster than its previous processor, dramatically increasing the accuracy of its decision-making.

Effectiveness, Commercial Promise, and Competition

While Google’s technology is more precise–it’s LIDAR system builds a 360-degree model that tracks obstacles better than Tesla, and can localize itself within 10 centimeters–Tesla’s is publicly available at a reasonable price.  Tesla’s most recent hardware set includes forward-facing radar, as well as eight cameras and twelve sensors around the vehicle.  The company continues to roll out new features in regular over-the-air updates.

To date, Tesla’s continuous push of new/updated Autopilot features has been (largely) successful in improving consumer safety.  Following a 2016 investigation into a deadly crash involving a Tesla Model S (which was closed without issue), the U.S. Department of Transportation found Tesla’s Autosteer feature had already improved Tesla’s exemplary safety record, reducing accidents by 40%, from 1.3 to 0.8 crashes per million miles.

 

Tesla’s software algorithms are a short-run competitive advantage over other automakers; its technology is in the hands of more users, quickly improving its solution.  However, as full-autonomous driving becomes commoditized over 10-30 years, the automotive business model will shift from vehicle ownership to transportation-as-a-service and the competitive advantage will shift towards mass-market fleet vehicle manufacturers (e.g., Toyota, Ford, GM).  If vehicles aren’t owned by the end-user, and, instead, summoned or rented, the need for a superior driving experience drastically decreases in favor of the cheapest fare. Accordingly, GM invested $500M in Lyft last year to begin building an integrated on-demand network of autonomous vehicles.

Improvement and Alterations

Tesla has made progress since its first software push, but according to Elon Musk–the company is multiple years away from pushing out Level 4 capabilities.  Moving forward, Tesla’s biggest obstacles (beyond regulation) are better local road mapping; removing the need for user input; and stronger recognition of stop signs, traffic lights, and road updates.  In most geographies, many Autopilot features are geoblocked, restricting use primarily to highways and other major roads.  By training its software to better recognize stop sign images, as well as traffic light locations and color changes, Autopilot can be utilized in more local situations.  In addition, Tesla’s publicly-available vehicles are not yet truly autonomous, even on highways.  Vehicles have hands-on warnings that require the driver to be engaged throughout the ride, as well as a feature that shuts off Autopilot for the remainder of the drive cycle if the driver fails to respond to alerts (“Autopilot strikeout”).

 

Tesla’s Autopilot In Action

Blog post by Ex Machina Learners

Sources

Ackerman Evan. “GM Starts Catching Up in Self-Driving Car Tech with $1 Billion Acquisition of Cruise Automation.” IEEE Spectrum: Technology, Engineering, and Science News. N.p., 14 Mar. 2016. Web. 07 Apr. 2017.

“Autopilot.” Tesla, Inc. Apr. 2017.

Fehrenbacher, Katie. “How Tesla’s Autopilot Learns.” How Tesla’s Autopilot Learns. Fortune, 19 Oct. 2015. Web. 07 Apr. 2017.

Habib, Kareem. “Automatic Vehicle Control Systems.” U.S. Department of Transportation NHTSA Announcement. Jan. 2017.

“NVIDIA CuDNN.” NVIDIA Developer. N.p., 30 Mar. 2017. Web. 07 Apr. 2017.

Pressman, Matt. “Inside NVIDIA’s New Self-Driving Supercomputer Powering Tesla’s Autopilot.” CleanTechnica. N.p., 25 Oct. 2016. Web. 07 Apr. 2017.

Randall, Tom. “Tesla’s Autopilot Vindicated With 40% Drop in Crashes.” Bloomberg.com. Bloomberg, 19 Jan. 2017. Web. 04 Apr. 2017.

Vijh, Rahul. “Autonomous Cars – Patents and Perspectives.” IPWatchdog.com | Patents & Patent Law. N.p., 06 Apr. 2016. Web. 07 Apr. 2017.