“‘How many quotes are you going to want that are going to be after this session?’” Larry Fink asked from the Davos stage. Elon Musk responded: “I don’t know. I mean, five.” But beneath the joking, the engineering bet in Musk’s message was clear: humanoid robots are ready to move from lab demos to real-world deployment, and Tesla’s Optimus is being positioned as the first mass-market contender once the reliability and safety hurdles are cleared. Musk said that Tesla already has its Optimus robots performing “simple tasks” in its factories and that more complex tasks in factories should be possible by year-end. Sales to the public would follow once “the level of reliability is very high, the safety is very high, and the range of functionality is broad,” he said.
The journey from a factory floor to a living room is not a marketing issue. It is a standards issue. Humanoids have industrial robot risks of pinch, crush, shear, entanglement, and impact with a distinct failure mode: legged mobility and falling. The standards community is still writing the rulebook, and this is important because it has implications for how issues are assessed. The work being done in ASTM’s legged robot systems subgroup suggests a test regimen that looks less like the robot cell checklists and more like stress testing balance recovery on uneven surfaces and moving platforms.
One of the implications is architectural, not aesthetic. Safety for humanoids requires redundancy that can trump high-level autonomy, and sensing that works even when the robot is in motion, with loads, and when it is being contacted. Reference frameworks such as ISO 13849 and ANSI/RIA remain relevant, but safety experts are also drawing on the discipline of vehicle safety, including ISO 26262 and ISO 21448, where consideration of edge cases and system-level hazard analysis is prominent. This blending of approaches helps to explain why “very high safety” is becoming a gatekeeper phrase, not a slogan.
Tesla’s own training method emphasizes that the critical limitation is operational learning, not merely mechanical design.
Within factories, Tesla has relied on massive data collection efforts to train Optimus on what to do and how to do it. At its Fremont factory, employees responsible for data collection have recorded themselves organizing parts and performing tasks along conveyors, which have been used to train the robot to replicate human actions. Tesla has indicated that it plans to start data collection at its Austin Gigafactory, targeting a start date of February, in a bid to increase the amount of real-world tasks and scenarios included in the data collection. The data collection system uses multi-camera helmets with a backpack system.
This approach is part of a larger trend in robotics, in which systems learn from “behavior” datasets in much the same way that language models learn from text. There has been talk of “Large Behavior Models” trained on videos of people performing physical tasks, with the goal of reducing the time it takes to train from months of programming to a fraction of that time. Other predictions have indicated that as much as 40% of household work could be automated in the coming decade.
Musk’s Davos presentation also linked robots to a less glamorous bottleneck: electricity. He said that power is the bottleneck in scaling AI, because the rate of semiconductor production is outpacing the rate at which new generations of power can be brought online. In this vision, lots of robots need lots of compute, and lots of compute need lots of power, and thus grid expansion, energy strategy, and high-density data center design become part of the humanoid roadmap. Meanwhile, researchers are also tracing the spread of automation in jobs.
A productivity model simulating the effects of generative AI predicted that exposure to the technology could increase productivity and GDP by 1.5% by 2035, with the greatest exposure in office and administrative task bundles. Meanwhile, legislatures are also acting on issues related to monitoring and managing in the workplace, further pushing the need for “human-in-the-loop” approaches.
The narrative that Musk wishes to convey is one of abundance: “essentially free or nearly free” AI, combined with robotics and an economic explosion “beyond all precedent.” The technical narrative that lies beneath is one of more limited and quantifiable standardized safety cases, robust fall and contact responses, scalable training data, and access to sufficient electricity to power the models that make humanoids useful outside of a controlled aisle.
