Figure 03 is Figure AI's third-generation humanoid robot. On 12 May 2026 the company put a team of Figure 03 robots on a livestream that has, at the time of writing, worked 119 hours straight.
The stream is still running. The title bar reads "Over 119 consecutive hours and 149K packages." Tens of thousands of people are watching humanoids flip parcels onto a conveyor, and the comment section underneath the video is, in its own way, more revealing than the demo itself.
What Figure actually showed
The official F.03 livestream went live on 12 May 2026 on Figure AI's YouTube channel. The framing is endurance: a team of Figure 03 humanoids running a continuous package-handling shift, broadcast in real time. Figure's own description claims the team has crossed 119 consecutive hours and 149,000 packages, working "at human performance levels" on the company's fully autonomous Helix-02 stack.
Three things worth saying about that framing. First, the numbers are company-stated. Nothing about the headline figure has been independently verified — there is no third-party auditor on the floor, no peer-reviewed paper, no published intervention log.
The stream shows duration and visible repetition. The livestream does not, on its own, show error rate, recovery rate, or the number of times a human in a Figure polo shirt walked into frame to nudge a stuck arm.
Second, "human performance levels" is Figure's wording. It belongs in inverted commas until somebody outside Figure runs the comparison. A separate YouTube viewer flagged the obvious experiment underneath the stream: in a head-to-head against a single human worker on similar tasks, the human kept full legal break times and still beat the robots by around 200 packages over the shift. The comparison is anecdotal, but it is the right kind of question to ask when a marketing line says "human-level."
Third, this is a stress test of the demo format, not just of the robots. The previous generation of humanoid showcases were short edited clips — a robot folding a single shirt, pouring a single drink. Figure is trying to retire that genre. The pitch is no longer "look at this one impressive motion." The pitch is "we can leave the camera on for a week."
The hardware behind the demo
ai/news/introducing-figure-03" target="_blank" rel="noopener noreferrer" style="color:#D97706;text-decoration:underline;text-underline-offset:3px">Helix, Figure's proprietary vision-language-action AI system. 02. Each hand now carries a palm camera.
The fingertips include tactile sensing capable of detecting forces as small as three grams. The vision stack is wider-field and lower-latency than the previous generation, and the robot can offload fleet-learning data over mmWave.
The detail that quietly does the most work for this livestream is the charging story. Figure 03 supports wireless inductive charging through a mat under the foot. The robot does not have to sit down to charge.
The livestream does not have to leave the line. It stands there, sips electricity, and keeps perceiving the world. The charging story marks the difference between a demo and a shift — the ability to top up without leaving the floor.
One viewer caught the consequence of that more clearly than the press release did:
Come to think of it, these robots have been standing the whole time — they charge while standing. No fatigue, no pain — that's the advantage.
— @modernetude5750, translated from Korean
The engineering story sits in one sentence. The Figure 03 advantage is not raw speed. It is the absence of fatigue, the absence of injury, and the absence of a lunch break. The hardware exists in service of a labour-pattern claim, not a dexterity claim.
The economics the comment section worked out in real time
The most useful piece of analysis underneath the livestream did not come from Figure. It came from a viewer doing arithmetic out loud. Here is the comment that got the engagement:
About 2.4kWh — not including charging-efficiency loss — per 3.5 hours. So, like, 10c an hour for labour. About $26k cost.
— @KenjiFox, on the F.03 livestream comment section
Run the maths he is doing. 5 hours puts the marginal electricity cost of running a Figure 03 at around ten cents an hour. Even on industrial rates and even adding charging inefficiencies, the figure is comfortably under fifty cents.
The hardware cost — KenjiFox eyeballs it at $26,000 — is a one-off. The labour cost, on this back-of-envelope, is electricity.
Ten cents an hour is the number Figure is selling without ever putting it on a slide. Ten cents is also the number the comment section instinctively reaches for. Once you set the comparison up as "twenty-six-thousand-dollar machine plus ten cents an hour versus a human warehouse worker on minimum wage plus benefits plus injury risk plus turnover," you have stopped arguing about robotics and started arguing about who gets the shift.
💡TK's take — on the economics
Figure 03 is, on this back-of-envelope, a labour-cost story dressed as a robotics story. The hardware buys the right to ignore lunch breaks, injuries, and turnover. The competitive question is not whether the robot can match a human on dexterity — Gartner-style operational benchmarks will eventually settle that. The question is whether warehouse operators choose the version that does not unionise.
The comment section is the real experiment
Look at any extended humanoid robot livestream and the same thing happens. People stop reacting to the robot. Viewers start reacting to what the robot implies. Within hours of going live, the Figure 03 comments had moved past "wow" and landed somewhere harder to summarise. A few examples, lifted directly from the F.03 stream chat:
One day humans will work for robots. Guarantee.
— @easy2do308
The "humans will work for robots" line shows up in some form every few minutes. The line is half joke, half resignation. The half-joke is comforting.
The resignation is the part that matters — viewers are not asking whether the labour shift will happen. They are quietly accepting the direction of the arrow and arguing about the slope.
They argue so hard for human rights but what about robot rights? Sweatshop boss is making it work 137 hours.
— @PoohJinping, translated from Korean
The "robot rights" line is the same instinct flipped. The viewer has watched a humanoid work a 119-hour-plus shift on camera and reached for the language of labour exploitation. The joke about robot unions follows in the next breath ("robots should form a union").
The viewer is not seriously proposing collective bargaining for embodied EI. The viewer is, however, importing the moral vocabulary of human labour onto a machine — and the import is happening without anyone asking whether the vocabulary fits. person Protocol.
When this gets its job done better and better this will mean human is a kind of animal that should be behind cages.
— @desteklisalla
Comments like that one are doing the dark version of the argument out loud. The comment is not a serious prediction. The comment is the inverse of the "humans working for robots" line, and it carries the same anxious logic — if the work goes to the machine, what does that say about the worth of the worker. The comment section is, in real time, rehearsing the cultural settlement that the technology will eventually force.
We are cooked, chat.
— @Raa1n
And then, near the end of the run, a different note — quieter, sadder, and the one I will not stop thinking about:
The fact that humans used to do this — I have real respect for those part-timers.
— @beggarleader, translated from Korean
The "real respect" comment is the entire piece in one line. The Figure 03 livestream is not, in the end, about whether the robot can flip a package the right way up. It is about a viewer watching a machine do something a person does, watching the machine do it for five days straight, and being moved to thank the person retroactively. The respect arrives only once the work has been visibly externalised.
What the livestream still does not answer
A stress test is not a benchmark. The 119-hour figure is real in the sense that the stream has been running for 119 hours; it is not real in the sense of being an independently audited measure of reliability. There is no published intervention rate.
There is no published failure rate. There is no comparison against a conventional warehouse-automation system or a human cell on the same SKU mix. The package-count headline is a Figure-stated number.
Several of the things a real buyer would ask are visible in the comments themselves. Why are the robots flipping the barcode away from the camera instead of leaving it scannable? Why does the same robot stumble on the same box shape three times in a row?
Why does another viewer count five red packages in quick succession, then three more a few seconds later — and how does that map onto the conveyor's actual throughput? 03 is ready for a real Amazon-grade shift. The livestream does not answer them.
What the livestream does establish is the new floor for the conversation. After this run, "humanoid robots cannot stay on task for long enough to do real logistics work" is a harder claim to make. The bar moves from "minutes" to "days." The argument shifts from feasibility to economics, ergonomics, and governance. Figure precisely wants the conversation to be having, because it is the conversation that ends in purchase orders.
💡TK's take — on the comment section
The viewers underneath this livestream are doing the cultural work the philosophers and policymakers are still scheduling meetings about. The category drift — labour vocabulary onto machines, dignity language onto code — happens in chat boxes first, white papers second. Read the comment section, then write the policy.
What I make of it
The Figure 03 livestream matters less as a research result and more as a confidence-building public test. The livestream is a piece of marketing engineered to retire one objection — "humanoids cannot work a shift" — by visibly working a shift. On that narrow goal, it is succeeding. The next round of objections has already shifted to electricity bills, throughput parity, and what it means for the warehouse workers whose jobs the demo is implicitly bidding for.
The harder, more interesting fact is the comment section. Tens of thousands of viewers are pricing the electricity, joking about robot unions, comparing to human throughput, and thanking the part-timers retrospectively. That public reasoning, happening live underneath a live demo, is the part of this moment we will look back on.
The robots are visible because the cameras are pointed at them. The settlement is happening underneath, in the chat.
The technology question is whether Figure 03 can do the work. The technology question will get answered, one way or the other, inside the next two years. The harder question — the one the comment section is already arguing about — is what the rest of us do with the answer.
Frequently Asked Questions
These are the questions readers have been asking since the F.03 livestream went up. Short answers follow, drawn from Figure AI's official product page, the livestream description, and the data in the comment section itself.
What is Figure 03?
In short, Figure 03 is Figure AI's third-generation humanoid robot, redesigned around the Helix vision-language-action AI system. The answer, simply put, is that Figure 03 is the robot Figure built to leave on a livestream for a week without intervention. The key is that Figure 03 is a labour-pattern claim wrapped in a hardware platform — wireless charging, palm cameras, three-gram fingertip sensing, mmWave fleet learning.
How does Figure 03 work with Helix?
Figure 03 accepts visual, language, and proprioceptive inputs through Helix, the company's vision-language-action stack. Research from Figure shows that Helix handles perception, grasp planning, and reorientation continuously over the 119-hour run. Data from the livestream description reveals 149,000 packages handled across the period, with the robots topping up via wireless charging mats without leaving the line.
Why is Figure 03 different from earlier humanoid demos?
Earlier humanoid demos were short edited clips. According to Figure's own product framing, Figure 03 is different because it was designed for near-continuous operation — wireless inductive charging, fleet learning over mmWave, and a sensory suite tuned for repetitive long-duration tasks. The answer is that Figure 03 treats endurance, not dexterity, as the headline benchmark.
Who is Figure 03 for?
Figure 03 is for warehouse and logistics operators, manufacturing customers, and — Figure says publicly — eventually home users. In other words, the platform is being positioned for any setting where the work is repetitive, the shift is long, and a wireless charging mat can sit under the worker. The commercial bet is on logistics first, with home use as the longer arc.
What are the real risks of humanoid robots in logistics?
03 livestream and Figure's own materials demonstrates three durable risks. First, the displacement risk — workforce cuts driven by the visible economics ("ten cents an hour for labour," as a viewer worked out). Second, the verification risk — company-stated uptime and throughput figures are not independently audited, and "human-level performance" is marketing language until a third party benchmarks it.
Third, the governance risk — evidence from the comment section shows viewers already importing labour-rights vocabulary onto machines, which is a category confusion the rules have not caught up with. Each risk is a policy problem, not a robotics problem.
Sources and further reading
