Figure 01 is being presented as more than a robot that follows narrow commands. In collaboration with OpenAI, robotics company Figure has built a system that can talk through a task, interpret what it sees, plan an action, and then perform it.
The key point is not only that the robot moves. It is that its movement is tied to a conversation and to visual understanding, so a request can be ambiguous and still lead to a useful action.
What Figure 01 can do
Figure calls the robot "Figure 01." According to the source, it connects to a multimodal model trained by OpenAI that understands both images and text. That connection lets the robot describe its environment, interpret everyday situations, and respond to requests that depend heavily on context.
The demo matters because the robot is not described as being remote-controlled. The actions in the video are learned, and they are executed at normal speed. That distinction is important: a remote-controlled robot can look capable on camera, but a robot using learned behavior is demonstrating a different kind of system.
Corey Lynch, robotics and AI engineer at Figure, framed the progress in striking terms:
"Even just a few years ago, I would have thought having a full conversation with a humanoid robot while it plans and carries out its own fully learned behaviors would be something we would have to wait decades to see. Obviously, a lot has changed."
Why the conversation matters
A robot that can talk is not new in the abstract. What is notable here is the connection between conversation, visual context, planning, and action. Figure 01 can use the discussion so far, including past images, to decide what a person means.
That matters because people rarely speak to machines in perfectly specified commands. A person might point, refer to something already mentioned, or use words whose meaning depends on the scene. The source gives the example of the question "Can you put that there?" In that case, the robot can use earlier parts of the conversation to infer what "that" and "put that there" refer to.
The same idea applies to everyday tasks. The source says the robot can understand that dishes lying around should probably go in the dish rack. That is a simple example, but it shows why context is central. The robot is not only matching a sentence to a fixed instruction; it is using visual input and conversation history to select an action.
How OpenAI's model fits into the system
The OpenAI-trained multimodal model is described as handling the robot's conversation and interpretation layer. It processes the full conversation history, including past images, and generates spoken responses that a human can answer.
The same model also decides which learned behavior the robot should perform to satisfy a command. In plain terms, the model helps bridge the gap between what a person says and what the robot should do next.
This is why Figure 01 is being described as a robot that can listen, plan, think, reason, and act. The source says Lynch describes capabilities that include describing visual experiences, planning future actions, reflecting on memories, and explaining conclusions verbally before or while actions are selected.
For a human user, that verbal explanation could make the robot easier to understand. If the robot can say why it is choosing a behavior, the interaction becomes less like operating a machine and more like working with a system that can account for its decisions.
The movement layer underneath
The physical actions are controlled by what the source calls visuomotor transformers. These translate images directly into actions, connecting the robot's camera input to movement.
The numbers in the source give a clearer view of the control loop. The visuomotor transformers process images from the robot's cameras at 10 Hz. They then generate actions with 24 degrees of freedom, including wrist positions and finger angles, at 200 Hz.
Those details help separate two parts of the system:
- Understanding and dialogue: the multimodal model processes images, text, conversation history, and spoken responses.
- Physical execution: learned robot behaviors and visuomotor transformers turn visual input into movement.
Together, those layers allow Figure 01 to respond to a spoken request with both an explanation and an action. The source does not describe this as a general-purpose household assistant. It presents a more specific achievement: a robot that can combine conversation, vision, planning, and learned movement in one demonstration.
Where this fits in robotics research
The source also points to similar work from Google with its RT models. Those models allow a robot to navigate an everyday environment and plan and execute complex actions based on the input and output of language and image models.
The difference emphasized in the source is that Google's demo robots were not as chatty. Figure 01 is being positioned around the combination of full conversation and action, rather than action planning alone.
That is the broader significance of the demonstration. Robotics progress is not only about stronger motors or better hands. It is also about whether a robot can understand what is happening, keep track of what has been said, choose from learned behaviors, and make its reasoning available through speech.
Figure 01 shows one version of that direction. It connects a robot body to an OpenAI multimodal model, uses learned behaviors rather than remote control, and performs tasks at normal speed. If the source's description holds, the result is a robot that can see, chat, plan, and act in a way that makes ambiguous human requests more usable.