Figure is trying to make household robots easier to command, and its latest step is Helix, a new machine learning model for humanoid robots. The company is presenting Helix as a way to connect what a robot sees, what a person says, and what the machine does next.
The announcement comes two weeks after Figure founder and CEO Brett Adcock announced the Bay Area robotics firm’s decision to step away from an OpenAI collaboration. With Helix, Figure is putting attention on a difficult goal: making a humanoid robot useful in messy, unpredictable home environments.
What Helix Is Designed To Do
Helix is described by Figure as a “generalist” Vision-Language-Action model. VLAs are a newer category in robotics that combine visual input with language commands so a robot can process its surroundings and act on a request.
The best-known example named in the source is Google DeepMind’s RT-2, which trains robots through a mix of video and large language models. Helix works in a similar direction, using visual data and language prompts to control a robot in real time.
Figure says Helix can handle household objects it has not seen before in training. The company writes, “Helix displays strong object generalization, being able to pick up thousands of novel household items with varying shapes, sizes, colors, and material properties never encountered before in training, simply by asking in natural language.”
That claim points to the main ambition: a person should be able to give a simple voice instruction, and the robot should inspect the scene, understand the object, and complete the task without a custom program for that exact situation.
Why Voice Commands Matter
For home robots, natural language is important because households do not operate like fixed production lines. A person might ask for a bag of cookies, a drawer might be open or closed, and objects may be placed differently every time.
Figure gives examples such as, “Hand the bag of cookies to the robot on your right” and “Receive the bag of cookies from the robot on your left and place it in the open drawer.” These examples are not just about one robot following one instruction. They involve two robots working together.
Helix is designed to control two robots at once, with one assisting the other on household tasks. That adds another layer of difficulty, because the system must coordinate visual understanding, language interpretation, movement, and cooperation between machines.
The broader idea is straightforward: a useful home robot cannot depend on a person selecting from a narrow menu of commands. It has to respond to ordinary instructions and adapt to the specific scene in front of it.
The Home Is A Hard Robotics Test
Figure is showing Helix through work with its 02 humanoid robot in home environments. That choice matters because houses are especially difficult places for robots.
Warehouses and factories are more structured. They often have repeatable layouts, predictable workflows, and controlled conditions. Homes are different. Kitchens, living rooms, and bathrooms vary widely, and the tools used for cooking and cleaning vary as well.
The source notes several household complications:
- People leave messes.
- Furniture gets rearranged.
- Lighting conditions change.
- Objects differ in shape, size, color, and material.
These conditions make learning and control major obstacles. They also help explain why home robots have not been the main focus for many humanoid robotics companies. The source points to five- to six-digit price tags as another reason the home has not taken precedence.
The usual approach is to work first with industrial clients, improve reliability, and reduce costs before moving seriously into dwellings. In that framing, housework is still a conversation for a few years from now.
Why Training Is So Costly
Figure argues that useful household robots need to create new behaviors on demand, especially for unfamiliar objects. The company says, “For robots to be useful in households, they will need to be capable of generating intelligent new behaviors on-demand, especially for objects they’ve never seen before.”
The current alternatives are expensive and slow. Figure says teaching even one new behavior requires “either hours of PhD-level expert manual programming or thousands of demonstrations.”
Manual programming does not fit the home very well because the number of possible situations is too large. A robot cannot realistically be hand-coded for every kitchen layout, cleaning tool, drawer position, lighting condition, and misplaced object.
Training through repetition has its own limits. Robotic arms in labs can learn pick-and-place tasks through repeated examples, but the source emphasizes the hidden work behind those demonstrations. It can take hundreds of hours of repetition to make a task robust enough for more variable conditions.
That is why object generalization is central to the Helix story. If a robot can reliably respond to new household objects through vision and natural language, the path toward useful home robotics becomes less dependent on manual work for each new task.
What This Announcement Signals
When TechCrunch toured Figure’s Bay Area offices in 2024, Adcock showed the humanoid being put through home-setting exercises. At the time, the work appeared not to be the main priority, as Figure was focused on workplace pilots with corporations like BMW.
The Helix announcement changes the signal. Figure is now making the home a priority in its own right, at least as a demanding environment for developing and testing training models.
That does not mean a household humanoid robot is ready for everyday buyers. The source is clear that work on Helix remains at a very early stage. It also warns that short, polished robotics videos can require a lot of behind-the-scenes effort.
The announcement is also described as, in essence, a recruiting tool meant to bring more engineers into the project. That framing is important. Helix is not just a product claim; it is a statement about the direction Figure wants to pursue.
If the model develops as intended, the key shift would be from robots that need carefully prepared instructions toward robots that can interpret ordinary household requests. For now, the practical takeaway is narrower but still significant: Figure is using the home as a proving ground for vision, language, action, and humanoid control working together.