How Tesla trains Optimus: camera helmets, human videos and 500,000 hours of data

A humanoid robot is only as useful as the skills it has learned. Building Optimus is a manufacturing problem, but teaching it to fold a shirt, sort parts or move a box is a data problem. Tesla's answer is a very Tesla one: lots of cameras, lots of video and one big neural network.

Step one: remote control

In its early years, Optimus learned mainly through teleoperation. A person wearing a motion-capture suit and a VR headset moved, and the robot copied those movements. The recordings then became training data. Tesla relied heavily on this approach until about June 2025.

Teleoperation also explains some of the criticism of Tesla's demos. At the "We, Robot" event in October 2024, observers noted that the robots interacting with guests were partly controlled by humans, and that Tesla was not open about it.

The method works, but it is slow. You need a person, a robot and a session for every recording. As Murtaza Dalal of Tesla's Optimus AI team put it in 2025, teleoperation "doesn't scale".

Step two: learning by watching

In May 2025, Tesla released a video of Optimus vacuuming, cooking and taking out the trash. The point was not the chores themselves but how the robot had learned them: from first-person videos of humans doing the same tasks. Milan Kovac, who then led the Optimus program, said Tesla could now transfer "a big chunk" of learning straight from human videos to the robots, with new skills triggered by voice or text and run by a single neural network.

From mid-2025, Tesla changed its data collection to match. Instead of motion-capture suits, workers wear helmets and backpacks fitted with five in-house cameras while doing routine jobs, such as folding a T-shirt or picking up objects. The footage trains Optimus to imitate those movements. It is the same camera-first logic Tesla uses for Full Self-Driving.

Step three: scale

Tesla has since spread the work across more locations:

• Fremont: the first training site, running for more than a year.
• Giga Texas: large-scale data collection started there in early 2026, with collectors filming tasks such as sorting vehicle parts and managing conveyor belts.
• Training hubs: according to The Information, as reported by Electrek, Tesla has added hubs in Colorado, Arizona and Florida, and moved much of its self-driving data labeling team to Optimus.

The same report put Tesla's training library at more than 500,000 hours, with a goal of doubling it by the end of 2026. Tesla still uses some motion-capture suits next to the camera rigs. Musk said back in January 2025 that Optimus would ultimately need at least ten times the training that the car needs.

The weak spot: generalising

Video makes it easier to collect lots of data, but it does not solve everything. According to people quoted by The Information in September 2026, Optimus still cannot reliably handle a wide range of tasks and can behave unpredictably in situations it was not trained for. Tesla is building a library of basic movements that the robot can combine, but teaching even simple tasks reportedly still takes several days.

Outside experts are split. Some think multi-camera video captures fine details such as finger and joint positions very precisely. Others argue that robots learn best by physically touching the world, which video cannot provide. Jonathan Aitken of the University of Sheffield summed up the approach as "a very Tesla way" of doing robotics, adding that no one else is trying it at this scale.

What comes next

• Third-person video: Tesla's longer-term goal is for Optimus to learn from ordinary videos of people, including clips from the internet, not just footage filmed through a helmet. That would open up an almost unlimited supply of training data.
• More compute on board: Tesla plans to add its AI5 chip to Optimus around April 2027.
• Proof in public: Musk has promised a video of Optimus doing tasks that were not pre-programmed. That would be the clearest sign yet that the video approach is paying off.