Will Tesla succeed in offering truly autonomous driving?

Zapping Autonews One-on-one with automotive design genius Henrik Fisker

The real surprise of this online show was a humanoid robot, called Optimus, which the company took the risk of presenting, although it is clearly not developed and cannot stand on its own. The reactions were also very mixed. However, this robot only occupied a third of the time of this session, which lasted a total of nearly three and a half hours.

Most of the presentation concerned the development of FSD (Full Self-Driving), the famous autonomous driving system which should be deployed by the end of the year worldwide. We were able to learn in passing that it was used by 160,000 customers and that no less than 35 updates had been made. Remember that the principle is to offer more advanced driving assistance in “beta” form than on the basic version of Autopilot, by adding recognition of traffic lights and signs.

Customers therefore become guinea pigs, and they pay for it ($15,000 since September 5, instead of $12,000 previously). A highly criticized process*. However, “all Tesla vehicles require active monitoring by the driver and are not autonomous”, can we read on the brand’s website. “With the FSD computer, we hope to achieve a higher level of autonomy thanks to the experience acquired by using our functions over billions of kilometres”, further specifies the American manufacturer.

*Ralph Nader, the U.S. attorney, calls this one of the most irresponsible and dangerous decisions an automaker has made in decades

The team behind the FSD

For this live, brand aficionados were able to discover engineers and developers in charge of the famous FSD, including Paril Jain, technical manager of Autopilot. Elon Musk had left his place for the occasion on stage.

Experts first explained how the system evolved, being able to make human-like decisions under critical conditions. For example, when a Tesla has to turn left in an intersection, it can now adopt a trajectory that avoids obstacles like a pedestrian who decides to cross. The system makes the synthesis between the geometry of the road (queues of traffic), the traffic and the presumed trajectory of pedestrians and vehicles. Artificial intelligence is based on a library of 4.8 million files combining text and images.

In one year, more than 75,000 cases were submitted to neural networks. That is, one every 8 minutes on average. The ultimate promise is to travel from door to door, respecting traffic lights and stop signs.

The next step is to build 3D models of the environment in less than 10 milliseconds from the video streams, in order to represent the possible obstacles that the car must manage. From the videos collected (Tesla is the only one based only on artificial vision), the brand wants to go even further.

Dojo, the super computer that processes videos

Photo credit – Tesla

She thus presented in detail the super computer on which she has been working since 2019. Called Dojo, it has a computing power of one exaflop (one billion billion floating point operations per second). This is double that developed by Fujitsu in Japan and which held the record so far.

Developed in-house, this computer is so powerful that during a test, it blacked out the power grid at a substation in Palo Alto, California. And the brand plans to acquire 7 copies. That’s why Tesla had to design a high-powered power system and another for cooling.

The Dojo will replace an existing supercomputer that used an Nvidia chip to process a 30 petabyte (30 million gigabyte) image bank. The computing power will make it possible to process the data generated by the Tesla fleet (3 million vehicles, which will cover 100 million km this year) and improve the Autopilot.

Elon Musk even thinks that his super computer could be used by third parties and be used to train artificial intelligence, and not necessarily only in the car for that matter.

A measured welcome

Of course, geeks find this awesome. And that probably appeals to the engineers, who were targeted through this event to encourage them to apply.

As for analysts, it’s not quite the same story. Thus, the American Gene Munster of the investment fund Loup Ventures, which follows the Tech, and Tesla in particular, considers that Musk is optimistic. And he warns: the FSD remains at a level 2, which is “light years away” from the required level 5, and it also requires the agreement of the regulator. Which will take years, he says. Similarly, the analyst doesn’t believe Tesla can match Amazon’s cloud capabilities for another decade or so.

During this time, the competition also advances and without bringing it back (nor taking lawsuits or investigations at the federal level). Thus, Aurora (start-up created by the former head of Google’s autonomous program) released its 4th generation of autopilot, Argo AI (subsidiary of VW and Ford) operates robot taxis on the Lyft network (competitor of ‘Uber) in Austin, Texas, and Waymo (subsidiary of Google) demonstrated the superiority of its technology compared to humans. Through a series of crash simulations, the artificial intelligence avoided more collisions than a very attentive driver (but not all of them), reports the company’s blog.

And we are not talking about the Chinese either, who are advancing very quickly on the subject.

to summarize

Our high-tech automotive specialist, Laurent Meilleud, talks in detail about the conference held by Tesla during which he presented his plans for the autonomous driving of his vehicles through his FSD computer.

Laurent Meillaud

Leave a Comment