What Happened
On May 13, Figure AI flipped on a public livestream from one of its test floors and let two humanoid robots get to work. Over the next two days, the machines ran a real package-sorting pipeline — grabbing parcels off a moving belt, reading labels, and dropping each one in the correct chute — without any human teleoperator stepping in. By the time the stream wound down, CEO Brett Adcock told Bloomberg the bots had logged roughly 50 hours of continuous operation and processed close to 60,000 packages, with one robot tagging out for a wireless top-up while the other kept the line moving.
The software stack on display was Figure's Helix-02, the second generation of its in-house humanoid foundation model. According to Adcock, the model ran fully onboard the robots — no cloud round-trip, no joystick operator off-camera, no human-in-the-loop scripting. The bots cycled through four-hour battery windows and self-summoned replacements when they detected a low charge, walking themselves off the line for a charging dock before another unit slid into the slot. Per-package time settled around three seconds, putting the robots roughly at parity with the human pickers the industry uses as a benchmark.
Some downstream reports went further. Seoul Economic Daily and a handful of Asia-Pacific outlets later put the total at 100,000 packages over 81 hours, citing a follow-up segment of the run. Either way, the message Figure wanted to send was the same: this was not a five-minute demo, and it was not a director's cut.
Why It Matters
For most of the humanoid robotics cycle, the genre's "wow" moments have been short. A backflip here, a tea-pouring demo there, a sizzle reel cut to dramatic music. What Figure attempted in this livestream is closer to what investors and operations leaders actually want to see: long, dull, uninterrupted utility. Package sorting is not glamorous, but it is one of the most measured tasks in modern logistics, with picks-per-hour and exception rates baked into every warehouse contract.
If the run holds up to outside scrutiny, it pulls humanoids one step closer to passing the only test that matters for a labor-replacing technology, which is whether a customer would buy them in volume. A robot that can match a human's three-second pick cycle, work through the night without a supervisor, and hand off to a peer when its battery runs low starts to look less like a research project and more like a piece of warehouse equipment with a depreciation schedule.
The broader pattern is also telling. The same week as the Figure stream, NVIDIA's Isaac platform announced new manufacturing partners, and physical-AI pilots that just a year ago would have been treated as long-horizon research kept clustering around real customer floors. Applied AI's center of gravity is moving from chat windows to conveyor belts.
Reaction
The response on X and tech forums split predictably. Investors and humanoid-robotics fans treated the livestream like appointment viewing — Bloomberg called it "Silicon Valley's latest binge-watch," and the embedded chat reportedly hit tens of thousands of concurrent viewers at peak. Brett Adcock leaned in, posting clips and statistics throughout the run and pinning a public commitment that no one would intervene, no matter what.
Skeptics pushed back. TechRadar and several robotics researchers flagged camera angles that conveniently obscured certain failure modes, asked why the run paused on cuts to other floor cameras, and noted that even with no teleoperation, the workflow itself had been carefully designed around what Helix-02 is known to handle. The most common critique was not that the robots were faked, but that the test environment had been groomed — a controlled sandbox, not a real customer's chaotic warehouse.
There were also softer concerns about what the demo represents. A widely shared post called the robots "stealing jobs from warehouse workers and streamers," which was half joke and half discomfort. The livestream worked precisely because it foregrounded the part of warehouse labor that most viewers had never thought to watch.
What's Next
Figure is now producing roughly 60 to 70 humanoid units per week at its main facility and targeting several thousand units annually. The company has previously disclosed pilots with at least one major automaker and one large logistics operator; if the livestream pulls more enterprise interest forward, expect Figure to publicize a new customer or a multi-site expansion in the coming weeks.
Helix-02 is also expected to keep iterating. The model is positioned as a single foundation model for humanoid control, with the same weights handling household chores in earlier demos and warehouse work in this one. Future updates are likely to widen the task envelope — irregular package shapes, unstructured shelving, mixed lighting — and to push more autonomy onto the robot itself, which is critical for customers worried about cloud latency or downtime.
On the policy side, regulators are starting to take notice. The European Commission's recent AI Omnibus deal extended deadlines for some high-risk AI categories, but physical-AI deployments in workplaces will still face scrutiny under occupational-safety frameworks and AI Act provisions on workplace monitoring. A 50-hour autonomous run is impressive; getting the same robot through an OSHA audit and a union conversation will be a different challenge.
Closing Thoughts
Demos like Figure's are useful precisely because they collapse the abstract debate about humanoid robotics into something more concrete. A camera, a belt, a robot, a number of hours, a number of packages. Either the bot does the work or it doesn't.
What this run does not settle is the question that actually drives the industry: cost. A humanoid robot that runs reliably for 50 hours is interesting; a humanoid robot that does so at a total cost of ownership beneath a human worker, across a full year of mixed shifts, is transformative. The next round of disclosures from Figure, Agility, and others will be much less about new tricks and much more about price-per-hour benchmarks and uptime guarantees.
For now, the takeaway is simple. Applied AI is becoming legible to people who do not read AI papers. They can watch a robot sort their next delivery in something close to real time, and they can decide for themselves whether the future of warehouse work is uncanny, useful, or both.
한글 요약
Figure AI가 5월 13일부터 약 50시간 동안 진행한 라이브 스트림에서 두 대의 휴머노이드 로봇이 사람 개입 없이 약 6만 개의 택배를 분류했다. 자체 개발한 휴머노이드 기반 모델 Helix-02가 로봇 내장 형태로 작동했고, 배터리가 떨어진 로봇은 스스로 충전 도크로 이동했으며 다른 로봇이 그 자리를 이어받았다. CEO 브렛 애드콕은 원격 조종이 전혀 없었다고 강조했고, 일부 매체는 후속 운영을 더해 81시간·10만 개 분류라는 수치까지 전했다.
이 시연이 중요한 이유는 화려한 묘기 대신 물류 현장에서 실제 평가받는 ‘긴 시간, 단조로운 작업’을 보여줬기 때문이다. 패키지당 약 3초의 작업 속도는 인간 작업자 수준이며, 야간 무인 운영과 자동 교대까지 결합되면 휴머노이드 로봇이 단순한 R&D 결과물이 아니라 ‘감가상각이 적용되는 창고 장비’로 평가될 가능성이 열린다. 같은 주에 NVIDIA Isaac 등 다른 물리 AI 파트너십이 잇따라 발표되면서, 응용 AI의 무게중심이 점차 채팅 인터페이스에서 컨베이어 벨트로 이동하고 있다는 분석도 나온다.
반응은 엇갈렸다. 투자자와 로봇 팬들은 ‘실리콘밸리의 새 정주행 콘텐츠’라며 환호한 반면, 일부 연구자들은 카메라 각도와 환경 통제 정도에 의문을 제기했다. Figure AI는 현재 주당 60~70대 수준의 휴머노이드 생산 능력을 확보하고 있으며, 연 수천 대 규모로의 확장과 차량·물류 분야 신규 고객 발표가 다음 변곡점이 될 전망이다. 결국 시장이 답해야 할 본질적인 질문은 ‘얼마나 신기한가’가 아니라 ‘시간당 비용이 사람보다 싼가’이다.