Figure Robot: Voice-Controlled Humanoid for Home Assistance

Figure Unveils Helix: A New AI Model for Humanoid Robots

Figure, a robotics firm led by founder and CEO Brett Adcock, has introduced a novel machine learning model designed for humanoid robots. This development arrives shortly after the company decided to discontinue a collaborative effort with OpenAI. The core of this announcement centers around Helix, a “generalist” Vision-Language-Action (VLA) model.

Understanding Vision-Language-Action Models

VLAs represent a recent advancement in robotics, utilizing both visual perception and language-based instructions to interpret and respond to information. Currently, Google DeepMind’s RT-2 stands as a prominent example within this category, employing a combination of video data and large language models (LLMs) for robot training.

Helix operates on a comparable principle, integrating visual input and language prompts to enable real-time robot control. Figure explains that Helix demonstrates a strong ability to generalize to new objects. It can manipulate thousands of previously unseen household items, differing in shape, size, color, and material, simply through natural language commands.

Bridging the Gap Between Vision and Language

The ultimate goal, according to Figure, is to allow users to simply instruct a robot to perform a task, and have it execute that instruction directly. Helix is engineered to connect the processes of visual understanding and language processing. Upon receiving a voice command expressed in natural language, the robot analyzes its surroundings visually before undertaking the requested action.

Examples provided by Figure include instructions such as, “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.” Notably, these scenarios involve cooperative interaction between two robots, as Helix is designed to manage and coordinate the actions of a pair of robots simultaneously.

Testing in the Home Environment

Figure is demonstrating the capabilities of its VLM through ongoing work with its 02 humanoid robot within a domestic setting. Homes present a unique challenge for robots, lacking the structured and predictable environments found in warehouses or factories.

Significant obstacles related to learning and control currently hinder the widespread adoption of complex robot systems in homes. These challenges, coupled with substantial price points, have led most humanoid robotics companies to prioritize industrial applications, aiming to improve reliability and reduce costs before entering the residential market. Household robotics remains a future consideration.

A Shift in Focus Towards Domestic Use

With the announcement of Helix, Figure signals a renewed emphasis on the home as a key area for development. This environment provides a demanding and intricate platform for testing and refining these training models. Successfully teaching robots to perform complex tasks in a kitchen, for instance, expands their potential applications across diverse settings.

“For robots to be truly useful in households, they must be able to generate intelligent new behaviors on demand, particularly for objects they have never encountered before,” states Figure. “Currently, teaching a robot even a single new behavior requires considerable human effort – either extensive manual programming by experts or thousands of demonstration examples.”

The Challenges of Scaling Robot Learning

Manual programming is not a scalable solution for home environments. The sheer variability of homes presents too many unknowns. Kitchens, living rooms, and bathrooms differ significantly, as do the tools used for cooking and cleaning. Furthermore, homes are often characterized by clutter, rearranged furniture, and varying lighting conditions. This approach is both time-consuming and expensive, though Figure possesses substantial financial resources.

An alternative is extensive training. Robotic arms frequently utilize this method for tasks like picking and placing objects in laboratory settings. However, achieving robust performance in variable tasks requires hundreds of hours of repetition. A robot needs to have performed an action hundreds of times to reliably execute it correctly the first time.

Early Stages and Future Development

Like much of the current work in humanoid robotics, the development of Helix remains in its early phases. It’s important to recognize that significant effort occurs behind the scenes to produce the concise, polished videos showcased. Today’s announcement functions, in part, as a recruitment tool, aiming to attract more engineers to contribute to the project’s growth.