Figure 04 vs Atlas: 4 questions that expose robot hype

Figure says Figure 04 is locked and parts are shipping. At its all-hands, it showed close-ups of parts, with no finished robot. Boston Dynamics went the other way and opened the workshop where Atlas gets built.

I put both companies side by side and sorted their numbers into three labels: measured, company-reported and target. My verdict: neither wins yet, and most of the headline numbers are promises or a company's own results. I only trust one result, a full shift with no human stepping in.

Four numbers that look like a league table

Four numbers from this year get ranked against each other all the time. They measure four different things:

• Build time: Boston Dynamics aims to build one Atlas in about an hour next year. Today it takes about a day.
• Line rate: Figure says its BotQ factory reached one Figure 03 per hour, in a post from April.
• Running total: Agibot announced its 20,000th robot. That counts robots built, with no information on how many are at work.
• Future capacity: Hyundai plans US capacity of up to 30,000 robots a year by 2028, in a factory it still has to build.

Figure's bet: data buys intelligence

Figure's new hand is on its fourth version, and Brett Adcock claims millimetre-level tracking against a human fingertip. That matters because a robot whose fingers land where yours land can learn from every video of a person working. Adcock called the tendon-driven hand on Figure 01 "one of the worst decisions" the company has made.

So Figure pays people to film everyday tasks. At the all-hands, Adcock said more than 210,000 people had uploaded video, from 112 countries, and that the data bill for the previous week was about $8.1 million.

The best evidence I found is Figure's Helix 2.5 test in 30 Bay Area homes the robots had never seen. Robots pretrained on human video finished the task 56% of the time, against 9% for robots trained from scratch. Each doubling of human video also cut test loss by a predictable step, across an 8x range.

I think this is the right bet, because adding data and compute is what made language models work. Those results are company-run though, and 8x is a short range. Nobody knows how far the line keeps falling for robots.

The 200-hour test

In May, Figure livestreamed one logistics job for 200 hours. Robots swapped out every few hours to charge, so no single robot ran 200 hours straight.

I still like this test. Two hundred hours tells me far more than a thirty-second clip, and anyone could watch it. Adcock also says Figure has robots at customers in three states. The only site I have seen on camera is BMW.

Boston Dynamics' bet: own the build chain

Boston Dynamics machines, assembles, repairs and tests Atlas in-house. Its team says an idea on Monday can be implemented by the end of the week. I've seen why that matters in my own work: a design that looks perfect on screen can fail under real load, and the faster you can make a part, the faster you learn.

The new Atlas hand goes from seven degrees of freedom to thirteen, and it has four fingers. A fifth would have needed more actuators, which means more cost and more parts that can break. To test the idea, the team taped their ring and pinky fingers together for a day and saw what they actually lost.

That tape test is the most useful idea in either presentation. Try the cheap version on yourself before you build the expensive one.

Atlas at Hyundai, and the plan

Atlas now trains at a new center at Hyundai Metaplant America near Savannah, Georgia, which opened in September. The job is part sequencing, putting parts in the order the line needs them.

Hyundai's plan is more than 25,000 Atlas robots across its plants, with sequencing from 2028 and assembly from 2030, the second date subject to technology validation. Boston Dynamics also targets a payback of under two years. I treat all of that as plans with dates. I count robots on shift, and those robots aren't on that shift yet.

The four questions I ask about any robot claim

1. Is it measured, company-reported or a target? 25,000 Atlas is a target. 56% is company-reported.
2. What does the number actually measure? Build time is different from output, and output is different from robots doing paid work.
3. What did the test allow? Figure's 200 hours allowed battery swaps. That's fine when the company says so.
4. Can I see it unedited? A robot shown only in part renders fails this one. A livestream passes.

My take

When I run both companies through these questions, Figure has the bigger data bet, and Boston Dynamics has the clearer path into a factory, even with Figure already at BMW. The winner will be whoever shows a full shift with no human stepping in, and a payback the customer measured.

I use the same test on AI agents. With agents, the demo was never the hard part. Running it every day was.