Figure announced Project Go-Big on September 18, 2025, describing it as a large-scale programme to collect data for humanoid-robot pretraining. The company also said its Helix vision-language-action model had learned to navigate cluttered home environments from human video, responding to natural-language instructions without robot demonstrations.
The result is a company-reported technical demonstration, not a peer-reviewed study or independently audited benchmark. Figure has not disclosed the number of video hours, homes, people or tested commands, nor a success rate or comparison with a robot-trained system.
Contents
- What Figure announced
- How the reported transfer works
- Why human video could matter
- What to watch next
What Figure announced
Project Go-Big is intended to build a large pretraining corpus for humanoid behaviour. Figure said it is working with Brookfield Asset Management to collect data in residential, office and logistics environments.
According to Figure, Brookfield's asset base includes more than 100,000 residential units, 500 million square feet of commercial office space and 160 million square feet of logistics space. Data collection had already begun in Brookfield environments and was expected to scale in the following months.
Those figures describe the environments available through the partnership. They do not establish how much data Figure has collected, how diverse the resulting dataset is or how representative it will be of the environments in which humanoid robots may eventually operate.
The announcement also extends the role Figure describes for Helix. The model had previously focused on upper-body manipulation tasks such as folding laundry, loading a dishwasher and reorienting packages. Figure now says one Helix network can produce both dexterous manipulation commands and navigation commands.
In the reported examples, a person can give instructions such as walking to a kitchen table or watering plants. Helix is described as navigating cluttered homes while generating control from visual input and language.
How the reported transfer works
Helix is a vision-language-action, or VLA, model. Such a model combines what a robot sees with a language instruction and produces actions or control outputs. In this case, Figure says Helix generates closed-loop control: it repeatedly uses updated visual observations to adjust its movement instead of executing one fixed route calculated in advance.
The training data is central to Figure's claim. The company says Helix was trained using 100% egocentric human video collected passively while people performed behaviours in real Brookfield homes. Egocentric video is recorded from a first-person viewpoint, allowing a system to observe the sequence of visual information and actions associated with a human goal.
Figure says the approach used no robot demonstrations and required no robot-specific data or training for the reported transfer. The intended mapping is from human-centred observations and behaviour to robot control: the model receives visual information and language, then produces navigation commands for the humanoid.
The reported navigation output is low-level planar motion control. Figure describes this as SE(2) velocity commands. SE(2) represents position and orientation in a two-dimensional plane, so these commands specify how the robot should move and turn while navigating.
That is different from simply identifying a destination. A system producing closed-loop SE(2) commands must continually convert its changing view of a room into movement decisions. It must also connect a language goal—such as reaching a table—with the visual conditions encountered along the way. The announcement does not provide enough architectural or training detail to determine how those steps are implemented.
Figure characterises the result as zero-shot human-to-robot transfer. In machine learning, zero-shot transfer generally means applying a capability to a target setting without task-specific examples from that setting. Here, the company says the humanoid learned from human video rather than demonstrations collected by the target robot.
Figure also calls the result the first time a humanoid robot has learned end-to-end from images and language to low-level SE(2) velocity commands using only human video.
Why human video could matter
Physical-robot demonstrations are costly and difficult to diversify across environments and tasks. That can limit the diversity of training data available for homes, layouts and behaviours.
Human video could offer a more scalable source of examples. People already perform navigation and household behaviours in many different spaces, and first-person recordings can capture the visual sequence associated with those behaviours. If a model can reliably transfer that information to a different body with different sensors and mechanics, developers may not need robot demonstrations for every environment or instruction.
That transfer is difficult, however. Human video records a human body, human viewpoint and human physical interactions. A humanoid robot has different camera placement, movement constraints, actuators and control signals. The model therefore has to learn more than a route: it must relate human behaviour to the robot's own available actions.
Combining navigation and manipulation in one model could also reduce the need to coordinate separate specialised systems. A household task often involves both movement and interaction—for example, navigating to an object and then handling it. Figure's announcement presents Helix as a single network that can produce both types of control, although it does not provide quantitative evidence showing whether this unified arrangement improves performance.
What to watch next
The most useful next evidence would be a quantitative evaluation with a defined test set, success criteria, failure cases and comparisons against robot-demonstration baselines. Results across different homes, lighting conditions, layouts, people and robot embodiments would help show whether the reported transfer generalises beyond the demonstrations.
It will also matter whether Project Go-Big improves manipulation as well as navigation. A larger corpus is valuable only if its scale and diversity translate into measurable capability, rather than simply increasing the amount of collected footage.
Finally, Figure's future disclosures may clarify how the system handles safety-critical situations, moving obstacles, ambiguous instructions and privacy in occupied environments. Until those details are available, the announcement supports an important research direction and a reported demonstration—not a verified general-purpose household robot capability.