Brett Adcock: Figure 6 months away from hours of autonomous work in any home?
Brett Adcock hopes Figure robots will perform hours of autonomous work in unfamiliar homes within six months. His conversation with Peter Diamandis reveals the strategy behind that ambition: Helix 3, redesigned Figure 4 hands, billion-dollar training runs, and factories that could eventually build robots without human labor. Consumer availability remains undated.
Figure 6 months from autonomous work in any home?
Brett Adcock thinks Figure AI could be less than six months away from putting a humanoid robot into a home it has never seen and having it do useful work for hours. As in, walk in the door, you tell it to vacuum, fry an egg, make coffee ...
You tell it what you want done .. the robot figures out how to do it.
That is the biggest claim in a recent, remarkably revealing conversation with Peter Diamandis on Moonshots. If Figure delivers, it would be a major step toward the general-purpose home robot people have been imagining for decades, and a much more consequential development than another carefully framed video of a robot completing a single chore.
“I obviously can’t promise, but I hope to see in the next sub-six months, perhaps even sooner, we can put a robot into any home and do fairly large, long horizon tasks end to end,” Adcock says.
The ambition is hours of work in an unfamiliar environment, directed through speech. And it comes with an important distinction: proving that capability is a separate milestone from selling a safe, reliable product that ordinary people can bring into their homes.
Figure has not announced that you can buy this robot six months from now.
But Adcock is putting a surprisingly short timeline on the intelligence breakthrough that would make such a product possible. Behind it are an unreleased Helix 3 model, a major redesign of Figure’s next robot, an enormous data collection effort, and plans for billion-dollar AI training runs.
I’ve covered much of the buildup, including Figure’s public all-hands disclosures at HumanoidMegaHub and the company’s broader humanoid strategy at Forbes. This conversation connects those investments to a much more tangible goal: a robot that enters your house and gets to work.
There’s a lot to unpack. There are also several places where the predictions run well ahead of the evidence.
End-to-end work in a zero-shot home
Adcock’s description of the target is unusually specific. Figure wants a robot to enter a home the company hasn’t trained in, encounter objects it hasn’t seen there before, take spoken instructions, and complete substantial work.
“My forecast, I think the first thing we need to prove to ourselves is we can do end-to-end work in a zero-shot environment,” he says. “I mean, a home we’ve never been in. We haven’t trained in it. We haven’t seen any objects before.”
Zero-shot is the key term. It means applying existing capabilities to a new situation without first training specifically for that situation. The robot has learned a lot about the physical world already; the question is whether that knowledge transfers to your dishwasher, your cupboards, your bed, and the particular chaos of your living room.
Adcock says Figure is already seeing this work at the level of individual tasks, such as making a bed or unloading a dishwasher. The next challenge is extending that capability over much longer periods.
That is a very different proposition from mastering one chore in a test kitchen. A household robot has to move between tasks, remember what it is doing, recognize when something has gone wrong, and recover without requiring its owner to become a part-time robotics technician.
Figure is testing in roughly 20 to 100 new California homes at any given time, Adcock says. Earlier Helix 2.5 evaluations involved 30 rental homes the robots had not trained in. Those are useful steps toward testing whether the system can generalize beyond familiar environments.
The six-month forecast pushes that much further. It suggests Figure believes the remaining distance between a successful individual task and hours of sustained work is now small enough to cross within a few model-development cycles.
That would be a substantial change in what humanoid robots are useful for.
Helix 3: the model behind the confidence
Before the conversation, the hosts saw an early Helix 3 checkpoint in Figure’s private “Black Box” test environment.
Adcock says they were likely the first outsiders to see it.
He describes the system as a two-year-old in an adult body: capable of attempting a wide range of requests, successful at some, and still struggling with others. The hosts describe seeing behavior that reminded them of a child experimenting with how to complete a task.
“And you can basically ask it whatever you want,” Adcock says. “And it can do a lot of things and some things it just can’t do well at.”
That is a more useful description than “baby AGI,” the phrase Adcock also uses. It tells us there is breadth, but also inconsistency. It suggests a model that can attempt unfamiliar work rather than one that is reliable enough to be left alone with an entire household’s to-do list.
Adcock says the checkpoint was significantly undertrained and that larger models were already being trained and evaluated. Later in the conversation, he says Figure has roughly 10 to 15 times more data sitting in its systems than went into the model the hosts saw.
The implication is that Figure sees a considerable amount of improvement still available through the training process it already understands.
There's more to learn here, of course ... the conversation does not provide a benchmark showing how often Helix 3 succeeds, how much assistance it needs, or how it performs over several uninterrupted hours. The hosts saw a private demonstration, and their impressions are encouraging, but obviously it'd be nice to see more.
The gap between doing a chore and being useful all day
One of the most important things Adcock says is also one of the least flashy: the robot needs to keep working.
“We’ve shown some of that, but we need to be able to do that end-to-end for a dozen-plus hours per day, every single day, seven days a week, for years,” he says.
That's the difference between a demonstration and a product.
A robot that folds a towel beautifully but needs help with every third shirt is creating a new supervisory job for human. A robot that unloads most of the dishwasher but gets confused by a bowl stuck to a plate still needs someone nearby. A robot that completes a task only when the room looks exactly as it did during training has limited value in a home that changes constantly.
Figure’s stated consumer standard is demanding.
“We don’t want to be teleoperating the home,” Adcock says. “We don’t want the robot to be having any issues and failures or stopping because the robot’s unsure what to do.”
He compares the intended experience to having another person in the house who can help. The robot should understand a request, act on it, and deal with the ordinary complications along the way.
That is the right target. The practical questions will be how close Figure gets, how the company measures success, and what happens when the robot encounters something outside its capabilities.
For the six-month milestone to be meaningful, I’d want to see task completion rates across many unfamiliar homes, human interventions per hour, and examples of robots recovering from mistakes. The failures will tell us almost as much as the successes.
Homes and factories are difficult in opposite ways
Adcock makes a useful distinction between Figure’s two markets. Commercial work is often easier to specify but harder to perform at the required speed and reliability. Household work can tolerate slower execution, but the variety is much greater.
In a warehouse, a customer can often describe the job precisely. Move these packages from here to there. Keep up with this conveyor. Work at this rate. The surroundings may vary, but the desired outcome is relatively clear.
The home is different.
“Every home is different,” Adcock says. “Things we’ll want the robot to do will change quite dynamically.”
Your appliances may differ from your neighbor’s. Your cupboards are organized differently. Your laundry includes different fabrics. You may want the kitchen cleaned one day and groceries put away the next. Even the same instruction can imply different actions depending on what has happened in the house that morning.
Figure is working on both markets in parallel because Adcock sees intelligence as the central constraint in each. Commercial customers need a system that can perform consistently under demanding conditions; consumers need one that can adapt to an enormous range of conditions.
Progress in one should help the other, but neither market automatically solves the other’s problems.
Figure’s big bet: physical intelligence improves with scale
The reason Adcock thinks the six-month target is plausible is that Figure believes it has found a workable way to keep improving robot intelligence.
“As we increased data, we saw a very log-linear relationship, meaning model losses were coming down, and the robots were getting better, actually executing better in these environments with more data than with less data,” he says.
This is the logic behind the spending. If additional data and compute reliably produce better physical performance, Figure can invest heavily in both with some confidence that the result will be a more useful robot.
It is a thesis I examined in my Forbes coverage of Figure’s recent company meeting: Adcock is betting on broad physical understanding developed during pretraining, rather than having to teach every deployment as a separate project.
The podcast adds a particularly interesting claim about transfer between tasks. Adcock says Figure collects examples of activities such as washing a dog or painting a house, and the resulting model becomes better at something else, such as folding laundry.
That does not mean a dog-washing video contains a secret laundry-folding instruction. It suggests the model may be learning more general patterns about bodies, objects, movement, and interaction that help across different tasks.
If the effect is strong and repeatable, it changes the economics of training. The value of a new example extends beyond the exact activity it depicts.
The missing information is the size of the effect, how it was measured, and how consistently it holds across tasks. Adcock describes a promising result, but the conversation does not include the experiments needed to independently assess it.
More data is only useful when it becomes better training data
Figure’s data operation is already enormous. As I reported in HumanoidMegaHub’s breakdown of the public all-hands, the company disclosed 37.5 million human videos collected through Index and an estimated $8.1 million spent on data in a single week.
The new conversation supplies an important qualification to those impressive totals.
“We have to make sure we clean it, annotate it, get it ready for training,” Adcock says. “It takes time.”
Figure also has to audit the data and determine whether it contains the right mix of examples. Then the training team has to incorporate it into a new run. A large backlog of video is potential value, not an immediate improvement in robot capability.
There’s a second challenge: the physical world contains information that ordinary internet video does not capture. Seeing someone pick up a cup does not tell you everything about grip force, touch, joint motion, or the feedback that lets a human avoid dropping or crushing it.
Adcock emphasizes that Figure needs rich information about physical interactions. The company is collecting from multiple sources, but the conversation does not fully explain how each source contributes to that physical understanding.
This is why raw hours of footage are an incomplete measure of a robotics dataset. Quality, diversity, relevant feedback, and the ability to turn the data into learning all matter.
Billion-dollar training runs are the next ambition
The other half of the intelligence strategy is compute, and the numbers get very large.
“I expect that next summer we’ll be doing billion-dollar training runs on Helix,” Adcock says. “If all goes well.”
Figure currently spends in the low nine figures annually on compute, according to Adcock. He says a single large pretraining run occupied more than half of its computing resources over the preceding month. And he makes clear that the company would use substantially more capacity if it could get it.
I covered the underlying infrastructure commitment in my Forbes report on Figure’s $3.5 billion Nscale deal. That agreement included an intention to scale beyond $6 billion, potentially using up to 100,000 Nvidia GPUs, with initial deployment targeted for the second half of 2027.
In this conversation, Adcock estimates that fully scaling the arrangement would provide well over 100 times Figure’s current compute scale. He again identifies the coming hardware as Vera Rubin.
There's lots going on here, but essentially, Figure is preparing to spend at a level that makes this much more than a competition to design a good electromechanical body. It is also a competition to finance and operate a very large AI development effort.
Compute availability is becoming part of the competition
Adcock’s account of buying computing capacity is revealing.
Earlier in the year, he says, Figure had several vendors competing for its business and negotiated favorable pricing. When the company went back seeking to double its capacity, the response was very different.
“We went out and got no bids back,” he says. Figure tried again, found a large cluster, and faced much less favorable pricing.
That is an account of Figure’s procurement experience, not proof that no compute is available anywhere. But it illustrates why advance commitments could become strategically important. A company can have the data, the model architecture, and the money, yet still wait for access to the infrastructure needed to run its next experiment.
Memory supply is another concern. Adcock says Figure has made partnerships to secure memory over the next few years and feels confident about its near-term ability to scale robot production.
Custom silicon is a possible longer-term answer, but there's a tradeoff: chip development takes years, while the models and their computing requirements are changing quickly. Designing too early around the wrong architecture would leave Figure with hardware that no longer fits the best way to run its intelligence.
Figure 4’s hands: two years of work and seven iterations
While the intelligence improves, Figure is preparing a substantial hardware change.
Adcock says Figure 4 will be the biggest generational step the company has made. The most striking detail concerns the hands: almost two years of development, seven iterations, and experience with both tendon-driven and non-tendon-driven designs.
“One of the biggest step-ups we’ve ever made in the hands is coming,” he says. He goes on to argue that a gloved Figure 4 and a gloved human would be “almost indistinguishable.”
That's a literally massive statement, so I look forward to seeing those hands in person. Humanlike appearance and motion do not establish human-level dexterity across a representative set of tasks. Force control, touch, speed, precision, durability, and the ability to manipulate awkward objects all matter.
But the importance of better hands is hard to overstate.
Almost everything a home robot needs to use was designed around our anatomy. Packaging, tools, cupboards, appliances, buttons, handles, and clothing all assume hands with a our fairly remarkable human combination of strength and sensitivity.
A significant improvement can therefore unlock many tasks at once. It can also make existing tasks less awkward and more reliable.
Adcock says the changes extend beyond the hands. Figure 4 will look substantially different from earlier models, and the company has taken some aggressive engineering bets to get there. Lower cost and greater capability are central requirements.
There is of course no release date, final price, or detailed specification sheet yet ...
Why the whole body matters as much as the hands
One of the most useful engineering lessons in the conversation comes from Figure’s earlier development.
Adcock describes the company’s coffee-making demonstration as a turning point. The robot’s upper body was controlled by a neural network while a conventional controller maintained its stance. It showed Figure a path toward learned manipulation.
Getting the robot to move around and work with its entire body took about another year.
The team initially underestimated the difficulty. It seemed plausible to move the robot to a location and then have it perform the manipulation task. But the useful placement of the hands depends on the legs, hips, torso, and balance throughout the movement.
“You need to understand what the knee and hip and feet are doing as it relates to fingers and hands,” Adcock says. “The same exact timestamp. So it needs to be like all trained together.”
Think about reaching into a deep cupboard or lifting a heavy box. We reposition our feet, shift our weight, lean, rotate, and adjust our grip as part of one coordinated action. Robot hands and arms in isolation don't see/know/feel/understand any of that, which is why a capable stationary robot arm does not automatically translate into a capable humanoid.
Essentially, putting manipulation on a mobile body adds a tightly connected control problem.
The robot needs to work when the internet doesn’t
Adcock says all of Helix currently runs onboard, including planning, reasoning, and low-level movement control. Real-time speech and fleet orchestration use servers.
“So, everything from the planning and the reasoning to even the low-level sending trajectories out and torques, all that’s done on the robot today,” he says.
The practical advantage is resilience. A robot should not stop carrying something or lose the ability to complete a physical task because the connection drops. Adcock says Figure’s robots can continue working when they lose internet access.
(This is a big argument against OpenAI's current bet against physical AI and a frontier model just running everything a robot does, by the way.)
That does not mean every feature works offline. Voice interaction and coordination are separate functions, and the company continues to experiment with the division between onboard and offboard computing.
Some control simply has to remain local because the timing is too demanding. The robot cannot wait through an unpredictable network delay while balancing or responding to contact.
Fleet management is different. Adcock describes coordinating robots as they leave a work area to charge, substitute for one another, or head to maintenance. Those are jobs where a shared system can organize multiple machines and their changing assignments.
A robot designed to help build more robots
Figure 4’s redesign also has a manufacturing objective. Adcock says the company changed aspects of the robot because Figure 3 would be difficult to produce through a fully automated process.
“We designed Figure 4 to help with this problem,” he says. “Part of this is, we think this is the robot that will build itself.”
The realistic interpretation is a manufacturing system in which humanoids operate machinery, handle components, and perform assembly alongside conventional automation. A humanoid does not have to personally manufacture every semiconductor or fabricate every piece of metal for such a factory to become highly automated.
It does, however, have to handle the tasks that still require people, including the exceptions that interrupt production. (I've seen this in operation already on Agile Robots' production line in Munich.)
Adcock says he feels confident that fully lights-out manufacturing can arrive within five years. He is much less definite about whether the actual interval will be six months, one year, or several years. Supporting that confidence, he reports robots climbing ladders repeatedly, carrying 50-pound boxes under neural-network control, and unloading items from a truck. Those capabilities extend the kinds of work Figure can attempt.
The current operating reality still includes substantial human support. As I reported previously, Figure is hiring robot babysitters for $45 an hour to coordinate robot operations, including overnight shifts.
More than 40,000 robots a year, and a path to a million
Adcock estimates that the existing BotQ facility could produce more than 40,000 robots annually across four lines at full rate. He also says Figure is working on additional manufacturing capacity, although the details have not been announced.
Beyond that, he sees a conditional path to a facility producing a million units per year. Getting there depends on solving the intelligence problems Figure is working on now.
The current internal fleet is already substantial. Adcock says there are well over a thousand robots on campus, approximately 1,500, compared with 700 to 800 people. Other robots are operating elsewhere.
Adcock says Figure also has several billion dollars in cash. His preference is to reach the general-robotics milestone while the company remains private, before the next phase of manufacturing, compute, and working-capital requirements.
The factory can only work as fast as the surrounding economy
Adcock makes another observation: a robot that can work around the clock may still be installed in a business that cannot.
“The truck doesn’t come at certain times,” he says. “It’s not like we can keep working. There’s literally no packages to move anymore if you’re working 24-7.”
That matters for both productivity forecasts and cost-per-hour calculations. Automating one task does not automatically automate the processes feeding it or the work that happens afterward.
A logistics line may still depend on human-operated loading, transport, maintenance, or receiving. A manufacturing cell may be waiting on components from a supplier that works a conventional schedule.
Adcock sees a later phase in which entire operations can run continuously. That is a much bigger systems problem than placing a humanoid at one workstation, but it is also where some of the largest potential gains would emerge.
A $300 to $500 monthly lease is the affordability target
For the home, Adcock says Figure will need something like a $300 to $500 monthly lease.
“In the home, we’ll have to make the robot really affordable,” he says. “You’ll need some sort of $300 to $500 a month lease.”
A lease could make sense when the hardware is improving rapidly. Customers may prefer access to a newer model every few years rather than owning an early generation for a decade. It also creates operational obligations. Someone has to maintain the machine, replace failed components, handle upgrades, and finance the gap between manufacturing cost and monthly payments.
Adcock says Figure intends to handle service itself initially and is working with large financial institutions on financing opportunities.
Safety requires judgment as well as reliable hardware
The household forecast also makes the safety discussion more consequential. A robot in an unfamiliar home encounters hazards that are difficult to enumerate in advance.
But there's physical safety as well as semantic safety:
“You don’t want a robot walking around and knocking over a candle or spilling the hot water,” he says. “These are things that you need to understand.”
A machine can execute a movement exactly as commanded and still cause harm if it misunderstands the situation. Conversely, a model can understand the task while a mechanical fault creates a dangerous failure.
Adcock points to more than 40 actuators onboard and the importance of avoiding falls, particularly around children and pets. He describes work across system architecture, low-level safeguards, cybersecurity, and repeated testing.
“And then you basically need to build a track record behind that,” he says.
That track record is, of course, essential. Interestingly, Adcock describes his young children wanting to jump on the robots ... and he had to tell that that they could not do that.
But kids will be kids ...
Your next robot could remember your last robot’s life
Another revealing part of the discussion concerns memory.
Adcock says Figure is developing robots that remember people and previous conversations. The intended home product includes companionship and familiarity as well as physical work.
The hosts ask what happens when someone replaces a robot after years of use. Does the personalization disappear with the old hardware?
“It definitely does not live with the device,” Adcock says. “You’ll be able to like have that pass through any device.”
Figure is also testing interactions in which someone walks past multiple robots while continuing what feels like a single conversation. Each machine can pick up the shared context. That would feel ... interesting.
That suggests a product in which the physical body is replaceable while the user’s preferences, history, and relationship with the system persist. Your next robot could already know where you keep the dishes and how you like your laundry folded.
Getting a new robot would be just like getting a new phone. Load it up from Apple's or Google's servers, and you're back in business with the new hardware ... and getting the Figure 05 or 09 would be just the same.
It could also become a powerful reason to remain with a particular provider: ecosystem effects. The practical value accumulates over time ... along with the sensitivity of the information the system holds.
China, Tesla, and the difference between a demonstration and a deployment
Adcock’s competitive commentary is pretty blunt. He says he has not seen much from Tesla recently that resembles the autonomous work Figure is pursuing, and he characterizes much of China’s visible humanoid activity as focused on dancing and gymnastics.
Asked to identify the closest competitor, however, he points to China collectively and acknowledges the pace of improvement.
“The amount of progress they’re making year to year is really great,” he says. “I would expect China to continue to do much better.”
His central technical distinction is useful: replaying an impressive movement and adapting to unfamiliar work are different achievements. A robot can demonstrate excellent balance and athleticism without being able to put groceries away in a strange kitchen.
The company is also making choices about where to compete. Adcock said (again) that Figure is pursuing civilian uses rather than defense applications. He says space-related discussions have occurred, but says (unsurprisingly) that Figure’s robots are not ready for space.
The immediate priorities remain work on Earth: commercial deployments, useful household autonomy, and the intelligence that could support both.
What would make the six-month claim real?
There are many potential headlines in this conversation:
- billion-dollar training runs
- a dramatically redesigned hand
- a million-unit factory ambition
- robots that may eventually participate in their own manufacture
The most important near-term test remains the unfamiliar home.
Adcock is suggesting that Figure is approaching a point where a robot can take an ordinary spoken request and perform substantial work for hours in a place it has never seen. That would move the conversation from whether a humanoid can learn individual chores toward whether one broadly capable system can become useful across many households.
Figure has identified the ingredients it believes will get it there:
- better general physical intelligence
- much more training data
- much more compute
- hardware built to support increasingly demanding work
Now it has also supplied a rough timeline.
In this story
Follow to get an alert when they're in the news
Humanoid (almost) Daily
The humanoid robot news that matters, with context from HumanoidMegaHub's tracking. Free, almost every day.
Keep reading
Written by
I tell stories with data and build startups. I've raised angel and VC funds, led software companies and startups, and chronicled the rise of the mobile economy. Currently I write for Forbes, host several podcasts, and am starting up HumanoidMegaHub. I was TUNE's mobile economist and Singular's VP of Insights, built the research team at VentureBeat, and managed teams creating software for partners like Intel and Disney. In past lives I built websites, e-learning platforms, and mobile apps, and, a long time ago, was named to Folio's top 100 of the media industry's "most innovative entrepreneurs and market shaker-uppers.”