Brett Adcock, the founder and CEO of Figure, unveiled a new machine learning model for humanoid robots on Thursday. This announcement comes just two weeks after Adcock revealed that the company would be ending its collaboration with OpenAI and instead focusing on in-house models. The new model, called Helix, is a “generalist” Vision-Language-Action (VLA) model that enables robots to process information using vision and language commands.

VLAs are a relatively new concept in the field of robotics, where robots use a combination of vision and language to understand and execute tasks. A notable example of this technology is Google DeepMind’s RT-2, which uses video and large language models to train robots. Similarly, Helix combines visual data and language prompts to control a robot in real-time, allowing it to pick up and manipulate various objects with ease.

According to Figure, Helix demonstrates strong object generalization capabilities, allowing it to pick up thousands of novel household items with varying shapes, sizes, colors, and material properties, simply by receiving a natural language prompt. This is made possible by the model’s ability to bridge the gap between vision and language processing, enabling the robot to visually assess its environment and perform tasks accordingly.

In an ideal scenario, users can simply instruct a robot to perform a task, and it will execute it without any issues. Helix is designed to make this a reality, allowing users to give voice commands to the robot, which then visually assesses its environment and performs the task. For instance, users can instruct the robot to “hand the bag of cookies to the robot on your right” or “receive the bag of cookies from the robot on your left and place it in the open drawer,” demonstrating the model’s ability to control two robots working together.

Figure is showcasing the capabilities of Helix by highlighting its work with the 02 humanoid robot in a home environment. The company is focusing on household tasks, which is a challenging and complex setting for robots due to the lack of structure and consistency. Houses are notoriously difficult for robots to navigate, with varying layouts, lighting, and objects, making it essential to develop robots that can adapt to these environments.

The development of complex robot systems for household use is hindered by difficulties in learning and control, as well as the high costs associated with these systems. Most humanoid robotics companies focus on building robots for industrial clients, which allows them to improve reliability and reduce costs before venturing into the household market. However, Figure is prioritizing the development of robots for household use, recognizing the potential benefits of having robots assist with daily tasks.

In 2024, TechCrunch had the opportunity to tour Figure’s Bay Area offices and witness the company’s work on its humanoid robot in a home setting. At the time, it seemed that the company was not prioritizing this work, as it was focusing on workplace pilots with corporations like BMW. However, with the announcement of Helix, it is clear that Figure is now prioritizing the development of robots for household use, recognizing the potential benefits and challenges associated with this market.

The announcement of Helix marks a significant shift in Figure’s priorities, as the company recognizes the importance of developing robots that can navigate and interact with complex household environments. By teaching robots to perform complex tasks in the kitchen and other areas of the home, Figure aims to create robots that can generate intelligent new behaviors on-demand, particularly for objects they have never encountered before.

According to Figure, for robots to be useful in households, they need to be capable of generating intelligent new behaviors on-demand, especially for objects they have never seen before. Currently, teaching robots new behaviors requires substantial human effort, either through manual programming or thousands of demonstrations. However, manual programming is not a scalable solution for household robots, as it is time-consuming and expensive.

The alternative is training, which involves repeating tasks hundreds of times to make a demo robust enough to take on highly variable tasks. However, this approach also has its limitations, as it requires significant time and resources. Figure recognizes that a more efficient solution is needed to develop robots that can adapt to the complexities of household environments.

While the work on Helix is still in its early stages, the announcement marks an important step towards developing robots that can navigate and interact with complex household environments. The company is using this announcement as a recruiting tool to attract more engineers to help grow the project, recognizing the potential benefits and challenges associated with developing robots for household use.