PHYSICAL AI

"Physical AI", broadly defined, is the application of artificial intelligence techniques to systems that physically interact with the real world. This is in sharp contrast to "Digital AI", where there are no real-world interactions. While both fields share many similarities in the underlying techniques used, the bar for safety with Physical AI is much higher.

Physical AI must reason against the range of human knowledge as well as against physics.

At Dexterity, we categorize Physical AI models into two types based on the pragmatic engineering behind them.

First, we have "Skill Models", which span sensory input, cognitive reasoning, and action. A limited view of these models would be a "Vision-Language-Action" model. A more expansive view with multi-modal sensory input and reasoning would be a "Sense-Think-Act" model. We believe that skills have the potential to closely model the human conception of motor skills—we have hundreds. Skills are likely recursive, which means they can be decomposed into sub-skills or composed into higher-level skills.

Second, we have "World Models", which provide a general understanding of how the world works. To gain a quick understanding of this idea, close your eyes and take a sip of coffee or tie your shoelaces. You'll realize that you rely upon a sophisticated understanding of your own body and the world around you, which is in addition to any motor skills you may have. This "World Model" serves as the foundation of our sophisticated dexterity, particularly as it pertains to multi-step tasks. Engineered correctly, world models should be capable of taking a short world history and predicting what comes next. This seems to be a foundational capability.

Ultimately, we believe it is too early to tag any single model as a "foundation model". But there are glimpses of foundational physical intelligence that are clearly evident in some AI architectures.

Our view on skill models comes down to two words: output interpretability. Skill models produce interpretable behavior. Notably, interpretability is a flexible criteria that does not implicate precision or accuracy. It simply states that if a skill is supposed to do something, like pick up a box and load it into a truck, then it does exactly that.

Dexterity's team is driven by real-world problems and our ability to have an impact. Real-world problems demand high performance in addition to interpretability. And the best way to deliver that is currently a fleet of smaller AI models. Engineered correctly, these smaller models can behave like building blocks and can be composed together to create more complex skills.

There will be thousands (maybe hundreds of thousands) of robots. If Physical AI isn't capable of generalizing, it will be limited in scope, will struggle to scale, and may ultimately fail to be commercially viable.

Classical manipulators are heavy, dangerous, require custom-programming to work effectively, and impose hefty CapEx costs and take up valuable space. They are extraordinarily useful for precision automation applications.

Humanoids are early and currently there's no widely adopted standard to certify safety.

Mech is deployed today to handle multiple workflows, at multiple locations, and is enterprise-grade.

We are always ready to engage with large enterprises. Whether you wish to be a customer today, or are just here to learn, our team of experts will be glad to engage.