“We are creating a new species”said Brett Adcock, Founder and CEO of Figure AI. He said this in a conversation with Marc Benioff (Salesforce) where they talked about humanoid robots. “We are creating a new species”.
It was at Dreamforce 2025. We saw this video again in 2026. And also, the image achieved some subsequent results.
Development of Figure 03
In particular, they showed the project called Figure 03 The company’s humanoids are the third generation. They are also a Artificial Intelligence called Helix (Around vision, language and action). The goal is for robots to perform tasks that humans do and learn from them.
There are several interesting things for the domestic sphere. On the one hand, they incorporate soft materials (textiles and foam) to avoid the risk of pinching, they have cameras on their palms and, in addition, they have touch sensors on their fingers. These cameras are useful when the main cameras are hidden (for example, when reaching for something in a closet).
Thus we have visual evidence of robots organizing houses or carrying out operations with dishes. It is important to know that just because a robot can do a task in a familiar environment does not mean it knows how to do it in a new environment.
We already commented on the deal between Image and BMW in our June 2025 coverage of Genbeta. In practical terms, you have to think about what they can do and under what circumstances.
In June 2026, BMW announced that it had Developing a project with Figure 03 in Spartanburg. The idea is that you collect the components (which come out of order) and put them in the car in the correct order for assembly. They also confirmed that Figure 02 was effective in producing more than 30,000 BMW X3s in 10 months. Specifically, they used robots to place (and then weld) pieces of sheet metal.
Test Helix 2.5 at home
On September 17, 2026 they introduced version 2.5 of Helix. They also tested in a home environment (30 homes in the San Francisco Bay Area). In this case, the system has not seen these houses before. They tried three types of tasks (collecting toys, folding towels and making the bed).
They reported a 56% success rate (with a pre-trained system indexed to its human behavior dataset), compared to 9% success with a model trained from 0. For it to count as a valid attempt, the task had to be completed completely.
The training data was collected elsewhere, but what was new were the houses and the evaluation objects. What we need to be clear about is that 56% doesn’t mean you can do more than half of any household chores; This is the test result.
Picture | YouTube with editing
