Figure 03 Wraps 200-Hour Autonomous Marathon At 249,560 Packages With Zero Hardware Failures — The Originally-Planned 8-Hour Livestream That Stretched 25× — Brett Adcock Pops Champagne At Sunnyvale HQ, Confirms 「Figure 04 Design Finalized」, May 22, 2026 | LostJobs.AI

The dashboard clock at Figure AI’s Sunnyvale demonstration warehouse rolled past 200:00:00 on Friday morning, May 22, 2026. The package counter read 249,560. The team popped a bottle of champagne behind the conveyor while a charcoal-grey Figure 03 wearing a name tag reading “ROSE” kept sorting.

Then Brett Adcock posted on X: “We just wrapped what began as an 8-hour challenge — and it ran for 200 hours without a failure. Shoutout to the team for the hardcore engineering behind F.03 and the robust Helix models powering it.”

200 hours. Twenty-five human eight-hour shifts. Zero bathroom breaks. Zero meal breaks. Zero hardware failures. Zero humans in the loop.

How an 8-hour challenge became 200

The marathon started as a dare. Industrial automation veteran Dr. Scott Walter issued a public challenge earlier in May: humanoids couldn’t run an 8-hour shift at human speeds without intervention. Adcock said: prove me wrong, we’ll do it live. The 8-hour livestream started on Wednesday, May 13.

Then the original “Bob, Frank, and Gary” trio crossed 24 hours at 28,000 packages. Then 60 hours. Somewhere along the way Aime, Figure’s intern, raced the robots for 10 hours and won by 192 packages — 12,924 to 12,732 — with a broken forearm by the end. Adcock’s response: “this is the last time a human will ever win.” The livestream kept running.

By the 200-hour mark on Friday morning, the fleet had been autonomously rotating between four robots — Bob, Frank, Gary, and Rose — for over 8 days. Per Sherwood News, the active robot swaps off the line every 3 to 4 hours when its battery runs low and walks to a wireless charging dock built into the floor; the next robot picks up the workflow without a human prompt. The Helix-02 neural network computes motor actions directly from raw camera pixels — no scripted behavior, no human teleoperator, no per-task code.

What “no failure” actually means at 200 hours

The marketing word “autonomous” carries a lot of weight in the humanoid space, so the specifics matter. Per Adcock’s own clarifications throughout the run, “no failure” here refers to two things: zero catastrophic hardware breakdowns across the fleet, and zero crashes of the core autonomy stack. The robots were not flawless on every package — viewers noted occasional dropped boxes and misoriented items — but those are sorting errors, not robot errors, and per Adcock’s framing, human workers drop boxes too.

The 2.6-second-per-package throughput floor that Figure posted in the first 8-hour window held steady across the 200-hour window. That is the part worth pausing on. A humanoid running open-loop for 200 hours typically degrades — actuator heat, wear in the wrists, drift in the perception stack. Figure 03 did not. The mechanical envelope held; the policy held; the battery-swap protocol held.

For automation veterans, this is the data point they have been waiting for since Brett Adcock first showed Figure 03 doing chores in October 2025. Two-second-class sort throughput, sustained over 8 days, with self-managed power, is the first credible public demonstration of a humanoid clearing the “shift work” bar.

The $39 billion bill

Figure is currently raising at a reported $39 billion valuation on a Series C. 200 hours and 249,560 packages is the slide that goes into that deck. It is a relatively narrow demonstration — boxes on a belt is not generalized AI, and no one in the building is pretending otherwise — but it is precisely the demonstration the analyst community has spent 24 months asking for. Can a humanoid hold its own across a workweek without a mechanical engineer babying it? Figure’s answer this week is yes.

The relevant industry comp for the demonstration is not other humanoid companies’ viral clips. It is third-party logistics operators. A modern parcel sortation operation runs three 8-hour shifts a day, 24-hour throughput targets, with about 5–10 minutes of unplanned downtime per shift considered normal. Figure 03 just demonstrated 25 of those shifts in a row with zero unplanned downtime. The benchmark has moved from “can a humanoid do this once?” to “can a humanoid do this cheaper than a sorter at a competitive wage?”

The cost gap is still huge in 2026. A Figure 03 at internal manufacturing cost is significantly above an annual fully-loaded sorter wage in the U.S., even before financing. But the depreciation argument — amortize the robot over five years of 24/7 shift work at a 99.99%+ uptime — is the argument Adcock has been building toward, and Friday’s number is the first one that lets him put it in front of an enterprise procurement team without a footnote.

Figure 04 is already designed

Tucked into Adcock’s wrap-up, per Sherwood: “the company said it has finalized the design of the next iteration of the robot, Figure 04.”

This is the part most readers will skip. It shouldn’t be skipped. Figure 03 was first publicly demonstrated in October 2025; Figure 04 design freeze in May 2026 implies a roughly 7-month iteration cycle for the next hardware generation, which is on the order of consumer-electronics timelines, not industrial-robot timelines. The capacity to iterate hardware that fast is the moat the closed-source humanoid camp is betting on against the Hugging Face / LeRobot open-source camp that shipped a $2,500 reference platform yesterday.

It’s also the part that should worry Hyundai/Boston Dynamics on their $25K Atlas at 30,000 units/year by 2028, Apptronik on Apollo, and 1X on NEO. A competitor that ships generational hardware every 7–9 months and just demonstrated a 25-shift-equivalent autonomous run is a very different threat shape than a competitor that ships every two years.

What this doesn’t fix

The package-sorting task is, as Figure’s own engineers would acknowledge, narrow. The robot is not doing dexterous unboxing, dual-arm wire harness assembly, or kitting variable-shape parts under deadline. It is, however, doing the exact thing that warehouse staffing agencies in the U.S. and Europe sell at $18–$22 an hour, 24/7, with attrition rates north of 100% annually. That market — third-party logistics sortation labor — is roughly 1.5 million workers in the U.S. alone. The labor-displacement math for that specific subset is now the part of the Figure investment thesis with public data behind it.

What the run does not yet show is task-generalization beyond sorting. There is no demonstration of Helix-02 transferring this kind of uptime to a different physical task — picking, packing, palletizing, unboxing — in the same livestream. That demonstration is the next bar, and Figure 04 hardware is likely the platform on which the company will try to clear it.

What to watch

Three signals over the next 60 days. First, whether Figure publishes a hardware utilization decomposition — how much actuator wear accumulated over the 200 hours, what the mean-time-between-recharge looked like, and what the per-robot duty cycle was. The 200-hour-zero-failure number is the headline; the wear profile is the procurement-team question. Second, whether a third-party 3PL operator signs a commercial pilot in Q3 2026, which is the milestone that turns the livestream from marketing to revenue. Third, whether Figure 04 ships before the end of 2026, which is what the design-freeze date implies but no one has formally committed to.

A 200-hour livestream of a humanoid sorting boxes is the kind of demonstration that would have read as science fiction in May 2024. In May 2026, it is the kind of demonstration that gets a champagne bottle popped at a $39 billion robotics company on a Friday morning, and a one-line confirmation that the next hardware generation is already through design review.