Figure AI did something that sounds silly but worked brilliantly: they put a robot sorting packages on a 24/7 livestream.
The result? It became one of the hottest things in tech this week.
What happened
On May 13, Figure AI started what was supposed to be an 8-hour robot demonstration. The star: the latest Figure 03 humanoid robot. The task: inspect barcodes on various small packages and place them on a conveyor belt with barcodes facing down.
CEO Brett Adcock initially set expectations low, saying a previous demo only lasted 1 hour and "high odds something breaks."
Eight hours passed. No failures. The team decided to keep going, 24/7.
By May 15, the robots had achieved 48 hours of nonstop autonomous operation without a failure. Adcock posted on X: "We are now running this until a failure to perform the use case."
Technical details
The robots run on Figure's self-developed Helix 02 neural network system, which allegedly enables full-body control and "long horizon autonomy." The system runs entirely onboard each robot's hardware—AI inference is done on-device, not in the cloud.
Training data: over 1,000 hours of human motion data, plus training across 200,000+ parallel simulation environments.
The robots are networked together—when one robot's battery runs low (each works about 3-4 hours), it can autonomously call another robot to take over. Hardware or software issues also trigger automatic swapping.
Why it went viral
Honestly, sorting packages isn't technically groundbreaking. Robotic arms in warehouses have been doing this for years.
But Figure AI did one thing right: they used a humanoid robot for it.
Humanoid robots have a natural "uncanny valley" appeal. YouTube commenters named the robots—Bob, Frank, Gary, Rose, Jim. People started betting on Polymarket about how long they'd last and how many packages they'd handle. Adcock himself appeared on camera wearing a T-shirt with "Frank's" image, simultaneously promoting the company's merch store.
This marketing approach is more effective than any technical whitepaper.
A moment of calm
But Ars Technica's Jeremy Hsu raised an important point: even the most impressive robot demos represent narrow windows for understanding real-world robot capabilities.
Sorting packages is a highly structured task—limited package types, fixed positions, repetitive motions. This is an entirely different difficulty level from having a robot cook in an unknown home environment or carry materials on a chaotic construction site.
Helix 02's "long horizon autonomy" performs well in the sorting scenario, but general-purpose capabilities still need more validation.
That said, the 48-hour failure-free number itself demonstrates progress in robot reliability. A year ago, demos that lasted 1 hour were considered successful.
Main sources: