Everything we learned about Figure from its public all-hands meeting

Public all-hands at Figure

Companies don’t usually publish their all-hands meetings. Figure just did.

On October 22, Figure CEO Brett Adcock posted the full meeting on X: an hour and one minute of slides, videos, inside jokes and a surprising amount of hard data on where the humanoid robot company is going. Figure’s roadmap has “completely changed over the last year,” Adcock wrote in the post, and the team has doubled. The meeting was his attempt to get everyone pointed the same way.

The meeting opened with an introduction from Arnold Schwarzenegger, kicking off the meeting with a bang. Then Adcock took over.

Here’s everything we learned.

Figure’s four chapters

Adcock’s framing for the whole company is a race in four chapters.

“We’re in a race,” he said. “We’re in a race to solve general robotics.”

Chapter one is building good robot hardware. Chapter two is building an AI recipe that lets that hardware do useful, autonomous work. Those two, he says, are deep engineering projects “that I think money can’t buy,” closer to orbital rockets than anything you can solve by hiring off the street.

Figure, Adcock argues, closed out chapter two this year.

Chapters three and four are scaling intelligence and scaling robots: making robots predictably smarter and then making a lot of them. 

“Thankfully, both of these problems are money can buy problems,” he said. “These last two chapters will … cost tens of billions of dollars.”

You can’t skip chapters, he says. And he took a clear shot at competitors who go straight to mass manufacturing before their robots are smart enough to be useful: “They’re building dead bodies that can’t run.”

From Figure 01 to Figure 04

Figure has now designed and shipped three generations of humanoid robots, and the fourth is coming.

Figure 01 was designed in 2022. “We had it up and walking in under a year since we filed the C Corp,” Adcock said. It was mostly an architecture test: actuators, battery, compute, sensors, kinematics and hands.

It also had a secret. When Figure abandoned its original tendon-driven hand, it needed new wrist motors and didn’t want to lose three months redesigning them. So the team borrowed actuators from the robot’s feet.

“So this robot actually has … forearms from the foot,” Adcock said, adding that “not a single person in like three years has ever said anything about it.”

Figure 02 was what Adcock calls “a feature complete robot.” The battery moved from a backpack into the torso, onboard compute doubled for Helix, the hands improved significantly, and the structure became more of an exoskeleton that carries loads, closer to how aircraft are built.

Figure 03 was designed for AI and for manufacturing. Cost dropped by more than 90% compared to Figure 02, mass came down, the sensor layout changed and the hands went through two more generations. Figure built its 1,000th Figure 03 over the summer.

Now it’s Figure 04.

“We are in full design lock, and we’re starting to ship parts,” Adcock said.

He says it’ll be the largest step up between generations Figure has ever made. A year and a half ago he thought Figure 03 had used up all of the company’s ideas. Now, he says, “I think we’re all out of ideas on Figure 04.” His comparison: “I think it’s the iPhone 1 version of humanoids.”

He also says he’s seeing every other new humanoid robot “going literally the opposite direction” from Figure 04.

The pieces are coming together in the lab. Figure 04 actuators are spinning in Figure’s dyno lab, and the company hoped to have every unique actuator fully characterized by the end of that week, after “pretty large sweeping changes” to their architecture. The week before the meeting, the team brought up 10 full Figure 04 battery packs for the first time, also with significant changes to both the cells and the pack. Full battery characterization, including thermals, comes over the next couple of months.

And the head, which Adcock says people told him was impossible, is apparently done too: “I heard the head was impossible in Figure 04, and look at that head like I saw last week.”

Of course: the hand is the hardest part

Adcock spent more time on the Figure 04 hand than any other component.

The tendon-driven hand on Figure 01, he now says, was “one of the worst decisions we’ve made to date.” Figure has since built multiple generations of hands, and the Figure 04 hand alone has gone through four major versions. Adcock describes it as “a system that is now more complex than an entire robot, has more actuators than an entire robot.”

The goal is human parity: millimeter-level tracking of human fingertips and contact points when grabbing objects. Every time the team found something a human hand could do that the robot hand couldn’t, Adcock says, “it basically just blew up the entire architecture of the hand,” because every millimeter of space is already spoken for by actuators, electronics, structure and wiring.

Version four is the one Figure plans to ship. “We think that is the hand that we’ll bring to market,” Adcock said.

Why the hands matter: learning from humans

The reason for all that pain is data.

Adcock describes what he thinks will become the biggest debate in robotics in the next few years, comparable to lidar versus cameras in self-driving cars. One path is designing specialized data-collection hardware and building the robot around it. That’s fast and fun, he says, and it has “the lowest ceiling.”

The other path is learning directly from humans, the largest ground-truth dataset on the planet. That means a robot body and hands that move enough like people for human video to transfer. It’s “the hardest path,” Adcock says. Figure is taking it.

He was just as blunt about companies collecting huge amounts of robot data at fixed tabletops with off-the-shelf arms: “If you collect millions of hours on a table, you will generalize to a table.”

Helix: leaving the table

Adcock walked through the three stages of Figure’s Helix AI so far.

Stage one was tabletop work, including the early Figure 01 coffee-making demo. When Adcock first saw it working, he wrote a number on a Post-it Note and gave it to the team: “I wrote five hundred million dollars.” His logic was that if they could pull it off, Figure could raise half a billion dollars.

“The team did it, and we raised a little under seven hundred million dollars,” he said.

Then Figure hit a wall. Moving from a table to walking around and manipulating things took a full year, because the team realized it was a whole-body problem. As a robot approaches a doorstep to drop off a box, its footsteps, posture and arms all have to respond to what’s in the camera frame.

“It’s not a manipulation problem. It’s not a tabletop problem. It’s a whole body problem,” Adcock said. “We had to re-architect Helix to be able to do this, and it took a year.”

Stage three is generalization: robots doing work in places they’ve never been. 

Last month’s Helix 2.5 release sent robots into 30 rented homes with different lighting, layouts and appliances, using objects Figure brought in to make sure they weren’t in the training data. Adcock says Figure learned three things: robots can generalize to unseen places, moving to pre-training works, and more data makes robots better.

Controls: stairs, ladders and molten lava

Adcock gave a big shout-out to Figure’s controls team.

Robots are now walking up and down stairs in homes they’ve never seen, with the goal of handling any stairs, commercial or residential. They’re also climbing ladders, and that’s a customer requirement. “The work is on a second story mezzanine,” Adcock said. “We have to go up and down to do the work. We’ll do it this year.”

Robots can now get all the way down to the ground and back up, use multi-contact control (putting a hand down to brace while reaching for something), stand on one foot under force and lift 25- to 50-pound boxes off the floor. Much of this used to be hand-coded. 

“These are things … a few years ago we were doing with thousands of lines of C++,” he said. “We can now do this with models.”

The controls team ships a new release roughly every month and is somewhere around version 14 to 17, Adcock says. A new controls-related announcement is coming next month: “we have some work we’re doing that will come out next month.”

(And yes, he showed a clip of a robot backflipping into molten lava.)

Then there’s the endurance testing. A Figure 03 on a treadmill was on roughly day 11 at the time of the meeting. And Bob, the robot from Figure’s logistics live stream, is still working. Figure set out to run the logistics task for eight hours, then just kept going.

“The robot ran over eight days, 200 hours straight, no breaks,” Adcock said. It swapped in a fresh robot whenever it needed a charge, and it’s “literally there right now,” still running every day. An intern raced it for eight hours and barely won, then showed up the next day with both arms in slings.

“I haven’t seen anybody do 10 minutes of this, let alone 200 hours,” Adcock said.

Robots are also getting out in public. 

“Mark Benioff sent me a text a few weeks before Dreamforce,” Adcock said. Benioff wanted robots there, and Figure had about two weeks. The robots walked through crowds autonomously, with no joystick, and talked to attendees.

Index: data at a scale robotics hasn’t seen

Robots get smarter with data. Figure’s problem is that the data it wants doesn’t exist.

“We went out and tried to buy data. It didn’t work,” Adcock said. “There’s no YouTube for robotics.”

So Figure built Index, the gig-style platform it launched in August that pays people around the world to film themselves doing everyday tasks: mowing lawns, chopping onions, driving forklifts. Figure doesn’t want a million hours from one factory. It wants the full distribution of everything humans do.

The numbers are big:

• 37.5 million videos uploaded, up from 16 million when I wrote about Index in early September
• 1.26 million app downloads
• 112 countries
• 210,000 unique people uploaded at least one video in the past week
• 20% average week-over-week growth for 10 weeks

“We’re printing eighty-one minutes of data every second,” Adcock said. At that rate, he says, Figure collects a full waking human lifetime every five days.

It isn’t cheap. 

“We spent … an estimated $8.1 million last week on data,” he said. That’s more than $420 million a year at last week’s pace, and growing. “We will spend more in data the next twelve months than most robotics companies have in total capital raise.”

Adcock says Figure is tracking about 50 startups trying to solve robot data collection. “Every week, we’re printing more than we could buy in the … entire market at this point.”

Collecting video is only half the job. 

Figure’s Helix infrastructure team cleans and annotates all of it, tracking what fingers and bodies are doing in space so it can transfer to robots. Human data, Adcock says, is “the largest untapped data source in the world.” Data is actually growing so fast that a recent finance review had him nervous about the projections: “Honestly, if anything, we need to kinda cool it off.”

Compute and scaling laws

Adcock wants Figure to be a “coin-operated machine”: put more money into data and compute, and get predictably smarter robots out.

The model is the 2020 scaling laws research for language models, which showed that more compute, data and parameters reliably reduce errors. Figure published what Adcock calls the first scaling laws for humanoid robot learning with Helix 2.5. As pre-training data increased, losses fell along a nearly perfect line, and robots scored better on structured evaluations.

“Wish the line wasn’t as straight,” Adcock joked.

That’s going to require a lot of compute. Figure’s deal with Nscale, which I covered in September, is an initial $3.5 billion commitment with intent to scale to $6 billion, running Nvidia’s Vera Rubin platform in West Texas. Figure is working through power permits and hardware procurement now, and the cluster comes online in 2027. In the meantime, Figure is doubling its existing compute within about a week and a half.

“We will need larger compute clusters than the LLMs have today over time,” Adcock said.

His reasoning: Figure may be “arguably the best post-training company in the world,” able to train a robot on almost any task it can get data for. But that isn’t enough. “Post-training is not gonna win this game,” he said. “It’s pre-training that will win.”

And pre-training is expensive. “This will take billions of dollars,” Adcock said. “You will not solve this with hundreds of millions or … a billion dollars.”

Manufacturing and deployments

Chapter four is building and shipping robots at scale, which means building a supply chain from scratch. Figure’s BotQ factory came up in about two or three months, Adcock says, and today Figure has more robots on campus than humans. More manufacturing capacity beyond the first BotQ facility is planned.

Commercially, Adcock was more guarded, saying customers have to sign off on anything Figure shares publicly. But Figure robots are deployed in three states right now, and named customers include JCPenney owner Catalyst Brands. One robot was recently dropped into a new customer facility, with different lighting and a different conveyor, and ran the job zero-shot with no new training. Another worked through a Los Angeles summer in a building with no air conditioning that hit 104 degrees.

Money

Figure had raised about $1.9 billion as of September. It has since taken in additional capital, Adcock said, without saying how much or from whom.

“We have over two billion dollar cash balance sheet today,” he said. More details will come “when the time comes.”

His ambition for all that cash isn’t small: “this will build the biggest business on the planet, and I want Figure to be that business.”

Baby AGI?

Adcock saved the biggest claim for last.

Last month, Figure trained a new model on its largest pre-training dataset and the most GPUs it has ever used, then tested it in a blacked-out house on campus the team calls Black Box.

“It was the most significant moment I’ve ever had in robotics,” he said. “I … think it’s baby AGI.”

He showed employees only a short teaser clip and said Figure isn’t ready to release it publicly. An even larger model is training now and should be ready in a few weeks, he says, and the models Figure trains over the next six months “are gonna be crazy.”

Baby AGI is a very big claim, and for now it’s just a claim. Figure will need to show the world what the model can actually do.

But Adcock isn’t planning to slow down.

“If we’re on the treadmill, we’re on the treadmill at nine point oh,” he said. “There is no slowing down now.”