Figure founder and CEO Brett Adcock revealed a new machine learning model for humanoid robots on Thursday. The news, which arrived two weeks after Adcock announced the Bay Area Robotics company’s decision to leave Openai collaboration, centers around Helix, a “generalist” vision language action (VLA) model.
VLA is a new phenomenon for robotics, leveraging vision and language commands to process information. Currently, the most well-known example of the category is Google Deepmind’s RT-2. It trains the robot through a combination of video models and major language models (LLMS).
Helix works in a similar way, combining visual data with language prompts to control the robot in real time. The illustration shows that “Helix displays the generalization of powerful objects and simply asks in natural language thousands of new household items that have never been encountered in training on various shapes, sizes, colors, and material properties. You can pick it up.”
In an ideal world, you just tell the robot to do something, and it’s just doing it. According to the diagram, it is where the helix comes in. The platform is designed to fill the gap between vision and language processing. After receiving a natural language voice prompt, the robot visually evaluates the environment and performs the task.
The illustration provides examples such as “Give the cookie bag to the robot on the right,” “Receive the cookie bag from the robot on the left and put it in an open drawer.” Both of these examples include a pair of robots to work with. This is because Helix is designed to control two robots at once, helping one to perform a variety of home tasks.
The figure introduces VLM by highlighting the work the company has done with 02 humanoid robots in a home environment. Considering its inconsistent with the structure of warehouses and factories, homes are notoriously tricky to robots.
The difficulty of learning and control is a major hurdle between complex robotic systems and the home. These issues, along with a 5-6 digit price tag, are why home robots are not prioritized by most humanoid robot companies. Generally speaking, this approach is to build robots for industrial clients, improving reliability and reducing costs before working on housing. Housework is a conversation in a few years.
When TechCrunch toured the figure’s Bay Area office in 2024, Adcock showed off some from the pace at which the company had passed humanoids in its home environment. At that point, this work seemed unprioritized.
With the Helix announcement on Thursday, the figure makes it clear that the house should be a priority in itself. This is a challenging and complex setup for testing this type of training model. For example, teaching a robot to perform complex tasks in the kitchen opens up a wide range of actions in a variety of settings.
“For robots to be useful in households, they need to be able to generate intelligent new behaviors on demand, especially for objects they have never seen before,” says the figure. “Teaching robots requires substantial human effort right now, even a single new action. Either doctoral level expert manual programming or thousands of demonstrations.”
Manual programming does not expand for the home. There are just too many unknowns. The kitchen, living room and bathroom differ dramatically from one side. The same goes for the tools used for cooking and cleaning. On top of that, people leave confusion, rearrange furniture and prefer different ambient lighting. This method is too time-consuming and expensive, but there are certainly many of the latter in the numbers.
Another option is training. Robotic arms trained to select and place objects in the lab often use this method. What you’re not seeing is that hundreds of hours of repetition are take-to-be robust enough to take on a highly variable task. To pick up something for the first time, the robot had to do so hundreds of times in the past.
Working with helix is still in its very early stages, like so many humanoid robotics at this point. Viewers are encouraged to create a short, well-produced video that will be seen in this post, where a lot of work is done behind the scenes. Today’s announcement is essentially a recruitment tool designed to feature more engineers and help your project grow.
Source link
