Tesla AI Director explains how vision-based strategies can make autonomy mainstream

It’s almost amusing, but rarely is Tesla’s Autopilot data considered when Wall Street analysts are evaluating the company. Even among its ardent bulls, very few apart from ARK Invest CEO Cathie Wood seems to take the company’s efforts at attaining full self-driving seriously in their forecasts. Part of this may be because autonomous driving is not really here yet. The regulations are not ready, and the technology is still catching up to the hype.

Tesla is dedicated to pursuing full self-driving using an approach that is heavily reliant on vision. Using its electric cars’ suite of cameras and artificial intelligence, Tesla can train its neural networks in how it responds to circumstances on the road. Other mainstream autonomous driving firms do not adopt this approach. Both Waymo and Cruise, both of whom are considered as the leaders in the field, rely on LiDAR, which Tesla CEO Elon Musk describes as a fool’s errand.

This approach was explained by Tesla AI Director Andrej Karpathy in a lecture last February, where he discussed the electric car maker’s strategies for its autonomous driving program. Karpathy covered several topics, from the challenges involved in long tail events as well as the importance of a vision-based approach in the widespread adoption and rollout of full self-driving solutions.

Critics of Tesla’s Autopilot and Full Self-Driving strategies would be quick to state that the company’s demonstrations, such as a clip of an autonomous Model 3 shown following Autonomy Day, were unremarkable since other companies like Waymo have done similar demos for years. This, Karpathy explained, is not the case, since Tesla follows a far different approach to achieve autonomy. “The critical point to make is that it looks the same, but under the hood, it’s completely different in terms of the approach that we take,” he said.

Companies like Waymo use LiDARs and high-definition maps to navigate around an area that’s intricately pre-mapped. It’s a sound approach, though it is also pretty restrictive. Waymo, for example, has been developing its autonomous fleet for years, but so far, it could only operate in a few locations. This is ultimately the reason why the grand scheme of things, a LiDAR-based FSD approach may not be feasible. There are simply far too many uncertainties on the road, both in highways and in inner-city streets, that would likely make a pre-mapped solution ineffective.

Tesla does not use high-definition maps. Instead, the company primarily depends on its vehicles’ cameras and artificial intelligence to drive without human input. This is very similar to the way humans drive, since people utilize their eyes to watch the road, and they use their brain to decide what actions to take while driving. Karpathy noted that this approach provides an opportunity for Tesla to roll out its Autopilot and Full Self-Driving improvements to its entire fleet efficiently. This allows Tesla to roll out an improved version of its Autopilot software that’s not in any way restricted to pre-mapped data.

Doing this is a complicated process, of course. To say that neural network training is a tedious process is an understatement, especially when it comes to long tail cases. Fortunately, this is a challenge that Tesla is all-too-willing to take. Karpathy’s lecture ultimately highlighted the foresight that goes behind Tesla’s efforts to develop a full self-driving solution that can be used anywhere.

It’s a bit amusing to note the similarities to the reactions that Tesla’s FSD efforts are receiving today and the reactions that the Model S program received from naysayers years ago. When the Model S was announced, critics noted that it was a fool’s errand. Others dismissed it as vaporware. No one has really made a premium, modern, production electric vehicle from the ground up then, so it would have been easier for Tesla to stay as a niche automaker. But staying as a niche automaker of high-performance electric sports cars doesn’t align with the company’s mission at all.

For Tesla to help accelerate the transition to sustainability, it must have mass market vehicles, and the Model S was the first of that. Doing so was difficult and almost like an act of hubris for the company then, as represented by the purchase of the Fremont factory that was incredibly large for the Model S production line. But Tesla did it anyway, and years after, the company stands just next to Toyota in terms of market cap. Similar to how Tesla eventually proved everyone wrong with the Model S, the electric car maker may very well end up proving its critics wrong once more when it releases its full self-driving features to its fleet.

Watch Tesla AI Director Andej Karpathy’s lecture in the video below.

The post Tesla AI Director explains how vision-based strategies can make autonomy mainstream appeared first on TESLARATI.