At an investor presentation on Monday, Tesla CEO Elon Musk laid out plans essentially to transition the company from a traditional car manufacturer to an operator of self-driving “robo-taxis.”

Musk reiterated a statement he made in January that all Tesla cars built since last month will have “feature complete” full self-driving capability by the end of this year. He said the company began building all cars with its new Full Self-Driving hardware computers starting with the Model S and Model X in March and all Model 3s starting in April.

The CEO presented Tesla as a whole new kind of car company, and Full Self Driving as its “killer app.”

He acknowledged that to enable full autonomy, where drivers can disengage and fall asleep—and eventually cars can be built without steering wheels or pedals—will require regulatory approval, but said he is confident that the company will have such cars operating in at least some locations by the end of next year.

READ THIS: Tesla starts Hardware 3 rollout: Enhanced Summon, red-light warnings

When Tesla announced the start of leasing for the Model 3 in early April, it specified that lessees no longer have the option to keep the cars at the end of the lease, but instead must return them to Tesla. Musk plans to use those cars to populate Tesla Network of robo-taxis after it takes them back. He said he expects to have the first Tesla Network robo-taxis operating in 2020.

Owners (who buy the cars and can keep them) will be able to share their cars on the Tesla Network to put them to work as robo-taxis when they don’t need them. Musk said the company will take a cut of 25 to 30 percent of the revenue, but estimated that for riders using the cars they could cost as little as 18 cents per mile. He estimated that a car on the Tesla Network will have up to three times the utilization rate of a non-shared, privately-owned car.

Musk also announced that, in support of the Tesla network, the company will begin engineering its cars and batteries to last 1 million miles, up from the current 300,000 to 500,000-mile target. He says the company’s Powerpack batteries are already engineered to last a similar number of cycles. If so, that would make Tesla the first company to do so.

“From now on, you’d be crazy to buy anything but a Tesla,” Musk said.

Tesla Model 3 dashboard in Autopilot testing with IIHS [CREDIT: IIHS]

Tesla Model 3 dashboard in Autopilot testing with IIHS [CREDIT: IIHS]

Musk says that its Autopilot system is improving exponentially as the company sells more cars and more drivers spend time using its Full Self-Driving and Enhanced Autopilot systems.

DON'T MISS: No more turn-signal affirmation for Tesla Navigate on Autopilot

Tesla buyers today can purchase the cars without paying for the Full Self-Driving Capability option; however Musk said that starting Monday, the company will begin to push the feature harder. Currently, it costs an additional $5,000, and the company says it will cost $7,000 if added after purchase. Full Self-Driving is a software feature that the company can send to cars over the air.

All cars built since 2016 have had the eight camera sensors, 12 ultrasonic sensors, and single forward radar required to operate Tesla’s Full Self-Driving system. The main Autopilot computer chip in older cars is not fast enough to provide full autonomy accurately—that is, safelythough cars back to 2016 are compatible with the new chip.

Tesla Model 3 all-wheel drive Performance rolls off a new assembly line in a temporary structure

Tesla Model 3 all-wheel drive Performance rolls off a new assembly line in a temporary structure

In a wide-ranging three-hour presentation, Tesla engineers and software developers dove into great detail about the new hardware chip, the learning “neural network” that Tesla has built to develop and improve its self-driving system, and the software and performance itself.

The new chip is 65 times faster than the old chip Tesla used, fits in the same space as the old chip (so that it can be retrofitted to some older models), and can process up to about 75 trillion operations per second, said Pete Bannon, Tesla’s lead Full Self-Driving hardware architect.

Andrej Karpathy, head of Tesla’s neural network, detailed how the company uses its fleet of 480,000 cars to gather data and video of unusual traffic incidents to learn how to improve the system’s performance. The cars have now covered 70 million miles using Tesla’s Navigate on Autopilot, according to Tesla.

Every time a driver has to intervene, a clip is sent to Tesla to analyze. For more complex traffic interactions, humans adjust the program to deal with the new situation. For more common interactions such as merging, the company has written artificial intelligence algorithms to analyze and correct the cars’ behavior.

In both cases, the software is fed back to the car via over-the-air-updates, analyzed in “shadow mode” (which doesn’t change the systems’ operation, but compares it to actual traffic, and eventually activated for a small group of “early access” users, if it improves the operation, then rolled out to the general population of Tesla drivers.

MUST READ: Tesla Full Self-Driving will still require drivers when it arrives later this year

With every iteration “we look at whether we’re getting closer and closer or farther and farther from how a human behaves,” said Stuart Bowers, vice president of engineering at Tesla. Among the examples he gave of difficult challenges the engineers have faced are cars carrying bicycles on the back, cars crossing multiple lanes and changing lanes simultaneously with another car. The goal is always to be able to anticipate where other road users are going, not just to identify where they are.

He said the company has developed images of what roads look like in bad weather, and expressed confidence that the system will be able to handle rain, snow, and darkness.

Musk doubled down on his expectation that expensive lidar sensors are an unnecessary dead end for autonomous vehicle development, saying that it only duplicates the data gathered from visible wavelength cameras and is less useful than radar, which can penetrate obstacles such as fog.

CHECK OUT: Tesla Autopilot ranks next-to-last in study of self-driving systems

He also dismissed systems such as that used by Google’s Waymo self-driving subsidiary, based on detailed mapping, because they can be fooled when anything in the environment changes slightly.

“If you have to have a geofence, you don’t have full self-driving,” he said.