Latest / Minds, Bodies, and Terawatts / Figure AI Humanoid Robots Achieve 24/7 Autonomous Work Milestone
Transcript
- 0:00[Host] This is Minds, Bodies, and Terawatts, where the robots never sleep and neither does the trash talk. Interesting Engineering reported today that Figure AI's Helix-02 humanoid robots completed round-the-clock autonomous operation, sorting packages nonstop without human intervention.
- 0:19[Guest] And the celebration was brief. Agility Robotics cofounder Jonathan Hurst told Business Insider at an event in San Francisco, quote: "Congratulations. We did that two years ago." Agility's Digit robots are already working shifts at Amazon, Schaeffler Group, and GXO.
- 0:41[Host] Fortune reported last year that Figure may have overstated what its robots actually did at BMW's Spartanburg plant. Adcock threatened legal action. Should we trust this demo?
- 0:55[Guest] The demo is real. Three robots sorted packages on a public livestream with viewers watching in real time. Helix-02 uses a unified neural network fusing vision, touch, and proprioception into one system, which is architecturally different from earlier approaches. But Hurst's point lands.
- 1:18[Host] Lands how? What is Agility doing that Figure isn't?
- 1:23[Guest] Paying customers running continuous industrial shifts. Figure just proved it can do that inside a controlled setting.
- 1:32[Host] So the milestone is hardware durability, not deployment readiness.
- 1:37[Guest] Right. A warehouse has pallets in the wrong spot, lighting that changes between shifts, packages that weigh different amounts every time. A controlled demo can run forever because the environment cooperates. The real test is whether Helix-02 handles the ten thousand things that go wrong on an actual floor.
- 2:03[Host] Which means the race is really about training data.
- 2:07[Guest] That's the arms race nobody's filming press releases about. Figure's tactile sensors detect forces down to three grams, the weight of a paperclip. But the neural network is only as good as the situations it trained on.
- 2:27[Host] And situations are what Agility has been collecting for two years.
- 2:33[Guest] Every paying customer becomes training data. Every dropped package, every weird pallet angle teaches the network something a livestream can't. Figure is starting that clock now.
- 2:47[Host] And Chinese firms aren't waiting. TrendForce projects 50,000 humanoid shipments this year, and Chinese manufacturers control roughly 87 percent of the market. Where does Figure fit?
- 3:01[Guest] The premium industrial tier. Their $39 billion valuation and the BMW pilot give them credibility with Western manufacturers who won't buy Chinese robots for geopolitical reasons. But Unitree ships its G1 at around thirteen thousand dollars. A factory can deploy ten Unitree units for what two Figure robots cost.
- 3:27[Host] Don't let Hurst's victory lap go unchecked either. Agility's Digit still operates behind safety barriers at its customer sites, physically separated from human workers. Working shifts and working alongside people are two different things.
- 3:45[Guest] The cooperatively safe Digit that can work next to a human isn't expected until late this year. Both companies are proving what their robots can do inside controlled zones while racing to handle the chaos outside them. Hurst's burn is also a sales pitch aimed at procurement teams: don't be impressed by a demo when we've been delivering production reliability.
- 4:17[Host] Livestream viewers named Figure's three robots Bob, Frank, and Gary. They proved the hardware does not quit. Now somebody has to prove it does not flinch. Patrick Déglon calls that gap between demo endurance and deployment reality the binding constraint on the Labor Cliff in Unscarcity.
- 4:41[Guest] For more on the data race behind humanoid robotics, read "How Do You Train a Robot" at unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.