Tesla Optimus Learns Chores by Watching Human Video

Tesla Optimus just got a new party trick: Tesla's AI team says the humanoid robot now picks up two-handed chores by watching people do them on video, instead of being puppeteered through every move.

What Tesla actually says changed

According to Startup Fortune, a single neural network trained on first-person footage of humans now lets Tesla Optimus vacuum a room, stir food in a pot, take out the trash and sort auto parts. Murtaza Dalal, a member of the Optimus AI team, put it bluntly: teleoperation does not scale, while training on video does.

Teleoperation means a human remotely pilots the robot's limbs to generate training data, one task at a time. It is slow, it is expensive, and it caps how fast any humanoid lab can teach its machines something new.

How the camera rigs work

Workers wear helmets and backpacks fitted with five in-house cameras and go about ordinary physical tasks, such as folding a shirt, loading a dishwasher or lifting an object off a shelf. That footage then trains the robot to copy the same motions, per the reporting Startup Fortune cites from driveteslacanada.ca and eWeek.

The approach replaced motion-capture suits and direct teleoperation after Tesla's head of Optimus left in mid twenty twenty-five. In a factory setting, the payoff looks concrete: the robot can identify a Model X fore link, pull it from a labeled box and set it on a dolly ramp, and the same network handles spoken or typed commands to switch jobs.

The bigger bet: raw internet video

Here is the part that matters most. Tesla is reportedly pushing toward training on generic, third-person internet video, the kind of footage never shot for robotics. That is unlabeled, filmed from random angles and effectively unlimited in volume.

It echoes the jump that took large language models from curated datasets to the open web. If a robot can translate someone else's camera angle into its own body movements, the way you can watch a friend tie a knot and then tie it yourself, the training pool goes from whatever workers filmed this month to every upload of hands doing something useful.

Elon Musk has already said on a January twenty twenty-five earnings call that Optimus training needs are probably at least ten times what self-driving required. Self-driving took roughly a decade of fleet data and still is not finished, so the scale problem is real.

The reality check

Do not picture a robot butler in your kitchen yet. Tesla has not published failure rates, task success percentages or how the system copes with tasks it has never seen demonstrated. And according to OptimusRumors, which summarized a report from The Information, Tesla is building hundreds of Optimus units a week, but most stay inside the company for testing and data collection.

That same summary says three people told The Information the software cannot yet handle a wide range of tasks reliably, can act unpredictably in new situations and still needs several days to learn basic jobs. Factory units reportedly stay in supervised areas and run specific programmed tasks.

There is also a talent story running underneath. Ashish Kumar, formerly the AI lead on the Optimus program, has co-founded a startup called Intelligent Machines that is betting on specialized industrial robots rather than humanoids, according to Crypto Briefing.

Why Gen Z should care

If video really can teach robots, the bottleneck in robotics shifts from hardware and human pilots to data. That could make machines cheaper to train and faster to spread into warehouses, factories and eventually homes, which is a big deal for the jobs you are about to apply for.

For now, the honest read is narrower and more interesting than the hype. Tesla believes it has found a mechanism to make the robot's learning curve look like a chatbot's instead of a decade-long slog. Whether Tesla Optimus can prove it outside a controlled demo is the thing to watch. For more on where machines are headed, browse our robotics coverage, and read the original at Startup Fortune and OptimusRumors.