Figure turns Brookfield's 100,000 homes into a recording studio for humanoid pretraining data
Figure AI said on 17 September 2025 that Brookfield, owner of more than 100,000 residential units, will help it build what it calls the world's largest humanoid pretraining dataset and will explore GPU data center and deployment work, and that Brookfield invested in its Series C. A day later Figure detailed Project Go-Big and said its Helix model had learned to navigate homes from first-person human video alone, without a single robot demonstration.
Figure AI has signed a strategic partnership with Brookfield that turns the investment group's apartments, offices and warehouses into recording sites for the human video Figure wants to use to pretrain its humanoid robots, the company said on 17 September 2025. Brookfield also invested in Figure's Series C as part of the broader arrangement, according to Figure's announcement, which did not give the amount.
Figure's release puts Brookfield's footprint at more than 100,000 residential units, 500 million square feet of office space and 160 million square feet of logistics space, held within a group that manages more than $1 trillion of assets. Figure said the partnership is meant to produce the world's largest and most diverse real-world humanoid pretraining dataset, to build AI infrastructure for scaling Helix, its in-house vision-language-action model, and to open new commercial settings for its robots. Neither side published financial terms, a target dataset size, the number of participating households or a date for the first robot deployment.
The agreement has three strands, according to Figure. Data collection had already begun inside Brookfield properties and would accelerate in the coming months. On infrastructure, the two companies will explore support for next-generation GPU data centers, real estate for robotic training environments and commercial use cases across Brookfield's global holdings. On deployment, they will study placing humanoids across Brookfield's portfolio over time. Figure added that it had started initial commercial deployments with some customers, without naming them in the release.
Brett Adcock, Figure's founder and chief executive, called the deal a major milestone and said Brookfield's scale gave Figure 'an unmatched platform' to capture humanlike navigation and manipulation data across household environments. Bruce Flatt, Brookfield's chief executive, presented it as part of the group's effort to use artificial intelligence to raise productivity in real assets and businesses. Stewart Upson, who leads global strategy at Brookfield, said the collaboration could unlock direct deployment opportunities over time, with a view to improving Brookfield's own operations.
On 18 September Figure published the technical counterpart, Project Go-Big, which it describes as its programme for internet-scale humanoid pretraining. The company argues that vision and language models advanced by training on vast corpora such as ImageNet, Wikipedia and YouTube, while robotics has never had a comparable store of physical behaviour. Because a humanoid's body resembles a person's, Figure says, ordinary first-person human video can be converted into robot training signal without the expensive teleoperated demonstrations, hand-written programs or staged lab settings the field has relied on.
The first result Figure reported concerns navigation. Using only egocentric video recorded passively as people carried out everyday tasks in Brookfield homes, the team trained Helix to follow spoken instructions such as 'Walk to the kitchen table' or 'Go water the plants', generating continuous control directly from camera images. Figure calls this zero-shot human-to-robot transfer and says it is the first time a humanoid has learned end to end, from images and language down to low-level SE(2) velocity commands, from human video alone. A single Helix network now outputs both dexterous manipulation and navigation, the company said. Earlier Helix demonstrations had focused on upper-body work such as folding laundry, loading a dishwasher and reorienting packages.
The partnership landed one day after Figure said its Series C had exceeded $1 billion at a $39 billion post-money valuation, an announcement dated 16 September 2025. Figure introduced Helix in February 2025, and its earlier announcements record a $675 million Series B in February 2024 and a $70 million Series A in May 2023. Three weeks after the Brookfield deal, on 9 October 2025, it introduced the Figure 03 robot.
The capital gap between Figure and other Western humanoid makers is wide in disclosed figures. Figure's Series C is the only round in the comparison below with a published post-money valuation, and the cash raised exceeds the latest disclosed rounds of Apptronik, Agility Robotics, 1X Technologies and Sanctuary AI. What none of those rivals has announced is a landlord partner offering recording access on the scale of Brookfield's portfolio, which is the asset Figure is now pricing into its story.
- Figure AI
- US
- No data
- Series C
- 1,000,000,000
- 39,000,000,000
- 2025-09-16
- Apptronik
- US
- 2016
- Series A extension
- 520,000,000
- No data
- 2026-02-11
- Agility Robotics
- US
- 2015
- Series B
- 150,000,000
- No data
- 2022-04-22
- 1X Technologies
- US
- 2014
- Series B
- 100,000,000
- No data
- 2024-01-13
- Sanctuary AI
- CA
- 2018
- Series A
- 58,500,000
- No data
- 2022-03-02
Latest priced venture round per company in the MEKAWIRE verified dataset. Amounts and valuations are as disclosed by the companies or their investors, not estimated; — means not disclosed. Figure's Series C amount is the disclosed floor (exceeded $1 billion). SPAC, IPO and undisclosed strategic investments are excluded.
As of Oct 3, 2026
The technique matters because data, not hardware, is the binding constraint for general-purpose robots. Teleoperation, where a person drives a robot to record a demonstration, yields precise training pairs but scales roughly with the number of robots and operators. Egocentric human video scales with the number of people wearing cameras, but it lacks the robot's own joint readings and must be mapped across what researchers call the embodiment gap. SE(2) commands describe motion on a flat floor, position in two dimensions plus heading, so Figure's claim covers walking to the right place rather than the harder problem of turning human hand motion into robot finger control.
For property owners the deal sketches a new role. Brookfield acts as data host, investor and potential customer at once, and its office and logistics space gives Figure environments that resemble the warehouses and back-of-house work where humanoids are most likely to earn revenue first. For other robot makers it raises the price of the data race, since access to real homes at this scale is hard to replicate through lab kitchens or rented apartments.
The open questions are material. Figure did not explain how residents consent to being filmed, who wears the cameras, how footage is anonymised or how many hours have been collected, and privacy rules for video recorded in homes differ sharply between the United States, Canada and Europe. The language of the infrastructure and deployment strands is exploratory. The navigation result is self-reported, without a public benchmark, success rates or a paper.
Three markers will show whether the partnership is more than a data licence. The first is a disclosed dataset size or hours figure for Go-Big, which Figure has not yet given. The second is a named Brookfield site running Figure robots commercially. The third is any Brookfield-backed GPU data center capacity dedicated to Helix training.
MEKAWIRE analysis
Figure is trying to turn real estate into a data moat, and the Brookfield deal is the first humanoid partnership in which access to human environments, not robot units, is the main asset being exchanged.
The evidence is in the structure. Brookfield's contribution is footage from more than 100,000 residential units and hundreds of millions of square feet of commercial space, plus an equity cheque and an option on infrastructure, while the robots it might eventually buy are left to future study. Figure's Go-Big result, navigation learned from human video with no robot data, is exactly the kind of evidence needed to argue that such footage is worth paying for.
The strongest counter-argument is that navigation is the easy half. Mobile robots have moved through buildings for years, and the commercial value of a humanoid lies in manipulation, where human hands and robot grippers differ far more than human walking and robot base motion. Until Figure shows manipulation skills learned mainly from Go-Big footage, the moat remains a hypothesis.
Bull case. If human video transfers to manipulation as well, Figure gains a training corpus whose cost scales with cameras rather than robots, and Brookfield becomes both its largest data partner and an anchor customer across apartments, offices and logistics sites. That combination would justify much of the $39 billion valuation on data advantage alone.
Bear case. Privacy constraints limit how much household video can be gathered and kept, rivals assemble similar corpora from public video and smaller landlords, and the infrastructure and deployment strands never move past exploration. The partnership would then read as a financing relationship with a marketing dataset attached.
Signals to watch:
- By 31 March 2026, the end of the 'coming months' window Figure set on 17 September 2025, a disclosed Go-Big dataset size in hours or households.
- Through 2026, the first Figure humanoid deployment at a named Brookfield property, with unit counts.
- Any Brookfield GPU data center capacity committed to Helix training, disclosed by either company, before Figure's next funding round.
- Egocentric (first-person) human video recorded in Brookfield homes
- Helix navigates from language commands using human video only, no robot demonstrations
- Dataset size, households, GPU capacity, deployment dates, financial terms
- Invested in Figure Series C, amount not disclosed
- 17 September 2025 (Figure newsroom)
- 500 million sq ft
- 18 September 2025 (Figure newsroom)
- 160 million sq ft
- More than 100,000
- More than $1 trillion
Why it matters
Humanoid developers have spent two years arguing over which data source will scale: teleoperation fleets, simulation, or human video. Figure's Brookfield deal is the first time one of the best-funded players has secured a large, contractually backed supply of real human environments for the third option, and it paired the announcement with a result claiming that human video alone can teach a humanoid to move through a home.
The deal also changes who sits at the table. A landlord with more than 100,000 residential units and a large logistics estate is now an equity holder, a data host and a prospective customer of a robot maker. That gives Figure a path from data collection to paid deployment inside one counterparty, which no rival has disclosed.
Rival analysis
Apptronik, Agility Robotics, 1X Technologies and Sanctuary AI have each built humanoid programmes on venture funding, and 1X has positioned itself most directly for homes. None of them had disclosed a property partner of Brookfield's scale at the time of the announcement. Their data strategies lean on teleoperation and customer pilots, which produce robot-native data but in far smaller volumes.
Chinese developers such as AgiBot, Galbot, UBTech and Unitree have leaned on volume production and deployments. Figure's bet is that household human video is cheaper per hour and more diverse than either approach. If that holds, Figure's rivals will need their own environment partners, which favours groups with ties to large landlords, hotel chains or retailers.
Valuation context
Figure's Series C, announced on 16 September 2025, exceeded $1 billion at a $39 billion post-money valuation, and Brookfield's participation was disclosed the next day without an amount. Among humanoid makers in the MEKAWIRE dataset, no other company has published a post-money valuation for its latest venture round, so a like-for-like valuation comparison is not possible from disclosed numbers.
The valuation embeds an assumption that Figure can scale robot intelligence faster than hardware rivals. Brookfield's data access is one of the few concrete assets supporting that assumption, which is why the dataset size and any manipulation results from Go-Big matter for how the next round is priced.
Supply-chain implications
Go-Big shifts part of Figure's input cost from robot hardware and operators to camera rigs, storage and GPU training capacity. The partnership's infrastructure strand, support for next-generation GPU data centers, speaks to the second of those, and Brookfield's infrastructure and power businesses are relevant suppliers of land, energy and buildings for compute.
For component suppliers the near-term effect is indirect. More training data does not change actuator or battery demand until deployments scale. The deployment strand inside Brookfield properties would be the trigger for hardware volume, and Figure has given no unit numbers or dates for it.
Signals to watch
The first signal is quantitative disclosure: hours of video, number of households or number of Brookfield properties in the Go-Big programme. Figure said collection would accelerate in the months after 17 September 2025, so an update by early 2026 would be consistent with that guidance.
The second is a manipulation result. Navigation transfer is the opening claim; a result showing that household manipulation skills can be learned mostly from human video would be the stronger proof. The third is commercial: a named Brookfield site with Figure robots under contract, and any GPU data center capacity tied to Helix.
Analyst view
Thesis: the partnership gives Figure a data supply advantage that is real but not yet proven to transfer to the tasks customers pay for. Confidence: medium. Brookfield's asset numbers, its equity participation and the start of data collection are confirmed by Figure's own pages, which supports the view that this is more than a memorandum.
Confidence is not higher because the technical result is self-reported, covers navigation only and comes without success rates, a benchmark or a paper, and because the infrastructure and deployment strands are framed as exploration. The privacy regime for household recording is also unaddressed in the public material.
Questions you should be asking
How do Brookfield residents consent to being recorded, who wears the cameras, and how is footage anonymised and retained? Does Brookfield receive any rights to the dataset or to models trained on it?
How many hours of video and how many households were behind the navigation result, and what success rate did Helix achieve? What share of Brookfield's investment in the Series C was linked to the data and infrastructure terms, and are there exclusivity provisions that would stop Brookfield hosting rival robot makers?