Tesla Autopilot’s vision neural net is ‘now working well’, but it needs to get real-world road data, says Elon Musk

As new owners of Tesla Model S and X equipped with the new Autopilot 2.0 hardware are growing impatient for not having the same convenience and safety features as owners of vehicles equipped with Tesla’s first generation Autopilot, CEO Elon Musk took to Twitter to reassure everyone that the company is making progress.

Tesla vehicles delivered since October 2016 are currently missing several convenience features like Traffic-aware cruise control, Autosteer, Auto lane change, Autopark, and Summon, as well as active safety features, like Side collision avoidance and Auto emergency braking. You can read the full list here.

From the release of the new hardware, Tesla always warned that the company would only release those features once they had been thoroughly validated to make sure they are as good as or better than the first generation Autopilot. From there, Tesla’s software team will then be able to improve on the features using the new hardware suite, Tesla’s in-house vision system (Tesla Vision), and its machine-learning technology with data collected from its fleet.

Today, Musk is now saying that Tesla Vision is up and running, but it needs to get some real-world road data:

Tesla Autopilot vision neural net now working well. Just need to get a lot of road time to validate in a wide range of environments.

— Elon Musk (@elonmusk) December 22, 2016

When introducing new features to its first generation Autopilot fleet, Tesla was using an extensive network of about 1,000 trusted owners in an ‘Early Access Program’. After Tesla goes through its own internal testing process and validates a software build, it pushes the update with instructions to those owners who then provide Tesla with feedback.

Since Tesla’s fleet of cars equipped with the new Autopilot hardware is much smaller, the company doesn’t have access to such an extensive test fleet, a fact that could be at play in the longer validation process since Tesla has to conduct almost everything internally.

It’s not clear how quickly Tesla can gather “a lot of road time” to complete its validation before pushing the update over-the-air to the fleet, but the company was previously guiding for “December 2016” (less than 2 weeks left) and a source familiar with the update told Electrek that they are still aiming for it.

Tesla releases new self-driving demonstration video with real-time ‘Tesla Vision’ feed

Tesla already has over 1.3 billion miles of road data accumulated through its first generation Autopilot program and that is useful for its second generation Autopilot, but the Autopilot team still also needs some direct data feeding Tesla Vision’s neural net with what the new hardware can see.

Once the update is released, Tesla will gradually push the features mentioned above to bring the new system to parity with the last generation and start improving from there with the new Autosteer+, Smart Summon, and other new ‘Enhanced Autopilot’ features.