Every viral humanoid-robot clip this year has arrived with the same shadow attached to it. When Figure AI posted footage in March of its Helix 02 robot tidying a living room, wiping a table with a spray bottle and tossing pillows onto a couch, Elon Musk fired back within hours with a two-word question: "Autonomously or remotely operated?" Figure founder Brett Adcock answered with two words of his own: "Fully autonomous." That exchange, more than any single demo, is the real state of humanoid robotics right now. The videos are extraordinary. The doubt is reasonable. And the gap between them is where the whole story sits.
The demos keep escalating. This week Adcock shared a clip of the Figure 03 robot climbing a ladder, a task that demands continuous balance and precise limb coordination, and one that AI researchers called genuine progress. In May the company ran an eight-hour live broadcast of a fleet of robots sorting packages in a warehouse without human intervention. Norway's 1X has shown its wheeled Eve robots folding shirts and wiping spills from a single learned model, a demonstration that pointedly echoed Tesla's earlier embarrassment, when a slick Optimus laundry-folding video turned out to have been teleoperated by a human off-camera. The pattern is clear: capability that looked like science fiction two years ago is now shipping as weekly content.
What is not shipping is verification. Figure has described its architecture in unusual technical detail, three stacked neural networks replacing more than 109,000 lines of hand-written control code, trained in simulation and transferred to hardware. That specificity is itself a kind of argument, because vague claims usually come with vague descriptions. But the company has not offered independent audits, raw sensor logs, or second camera angles for its headline demos. When Adcock says "fully autonomous," the industry is asked to take his word for it. After the Tesla laundry episode, that is a heavy lift, and Musk's question landed precisely because he has fielded the same accusation about Optimus for years.
Honda offers an instructive counterpoint, because it refuses to pretend. Its post-ASIMO strategy centres on an avatar robot that is explicitly remote-controlled: a human operator decides what to do while onboard AI corrects grip, force and hand position across a video link. Honda's new hand can thread a needle and tighten small fasteners, but the company is candid that a person stays in the loop. That honesty reframes the whole debate. The interesting question is not whether a robot can perform a task on camera, but how much of the intelligence is in the machine and how much is in the human you cannot see.
The stakes behind the semantics are enormous. Figure has raised roughly $1.75 billion at a $39 billion valuation, backed by Nvidia, Microsoft and Amazon, on the strength of demonstrations rather than disclosed revenue. Yet in actual deployments today, Chinese firms lead: AgiBot and Unitree together account for well over half the humanoid robots in the field, while Figure has not shipped commercial units at scale and its flagship has no announced release date. A viral video can settle the autonomy argument in a news cycle. It cannot settle whether an eye-watering valuation buys enough runway to turn a great demo into a product before cheaper rivals flood the market. Until the field agrees on how to prove autonomy rather than assert it, every stunning clip will keep arriving with the same two words attached.