Archetype AI is building a model intended to understand the physical world the way language models understand text.
Instead of training on digital artifacts, the company’s foundation model (called Newton) learns directly from sensor data. It ingests vibration readings, radar signals, inertial measurements, environmental data, audio, and video, and maps them into a shared representation that can be queried, monitored, or used to drive automated agents.
The Palo Alto-based startup recently raised a $35 million Series A to expand this platform and scale enterprise deployments. Its founding team came out of Google ATAP, where they built sensing technologies such as Project Soli, a pico-radar system later integrated into consumer hardware.
In an interview with AIM Media House, cofounder and chief scientist Jaime Lien points to years of work interpreting complex sensor signals as their foundation. “We focused primarily on understanding complex sensor data… being able to take this complex non-visual type of sensor data and interpret it into human-understandable forms of meaning,” she said.
The team saw a chance to generalize that work: “With the advent of large models unlocking things like self-supervised learning and generalization across many different modalities, we saw an opportunity to take some of the thoughts and applications that we had seen at Google and pursue it outside of that large corporation.”
Modeling the Physical World
Newton is designed to learn from the dynamics of physical systems. Instead of receiving human-defined equations or curated labels, it observes how measurements change over time and space. “By observing through direct measurements the way that physical processes and behaviors evolve, the model can learn underlying behaviors. Essentially artifacts of physics,” Lien said.
This approach places Newton within the broader conversation around world-model architectures. Lien drew a clear link. “This is analogous to these same concepts of world models [that] people are talking about,” she said. “Whether we call this a world model or not is semantics.”
Archetype’s interpretation extends beyond the camera-centric work that dominates the field. “A lot of these other world models that people are pursuing are very vision-based,” she said. “What we’re really looking into is this long tail of different sense modalities… things like electrical measurements, pressure, environmental, gas.”
Archetype’s technical architecture centers on a single embedding space. “We developed one common embedding space, so a common latent representation of all this data, whether it’s coming from an IMU or from language or from various other types of sensor modalities,” Lien said. From this space, Newton applies a shared decoding architecture. “We can then apply a common decoder… in order to extract meaning from that.”
Lien noted that the company has begun probing this latent space and is seeing early indications of learned structure. “We actually have some early results that are showing that indeed there is an encoding, a representation that captures the structure of the physical world,” she said.
A version of the model called Newton TimeFusion merges sensor data and natural language into a unified token space. This allows operators to query sensor streams using plain English, retrieving summaries and interpretations generated directly from sensor measurements.
Lien said this grounding in measurement, rather than text, has an additional benefit in operational settings. “Physical AI provides a grounding of large models,” she said. “Our models are trained on direct observations… and training these models to really ground on the sensor data and not over rely on textual prompts is one way to mitigate hallucination.” She added that physical domains often contain objective outcomes: “With physical applications, there’s often a hard ground truth… There’s a very clear right or wrong answer.”
Applied Physical AI
Newton is not delivered directly to customers. Instead, Archetype provides “Physical Agents,” which package the model’s capabilities into configurable applications. “The foundation model is really the underlying intelligence that powers all of these agents and all of these applications,” Lien said.
One early agent is built for industrial process monitoring. It can analyze vibration, temperature, and other machine-level signals to identify anomalies or transitions in operational state. “An industrial process-monitoring agent can take in different sensor modalities and monitor them for either anomalous behavior or track the different states the machine is evolving through over the course of time,” she said.
Another agent focuses on task validation. “These are agents that can take in a prescription of the way that a work task needs to be performed and then check whether the performance is actually matching what was prescribed.”
Agents can be customized with natural-language instructions. “English is the main programming language,” Lien said. “Anything that you can specify through natural-language type of instructions really opens up a wide array of different ways that you can customize.”
Archetype is also looking to help customers prototype new agents. “We’re building what we call the agent toolkit, which will provide simplified graphical interfaces for quickly prototyping and experimenting with agent configurations,” Lien said.
Early Deployments
Archetype’s early deployments span manufacturing, construction, and public infrastructure. The company partnered with the City of Bellevue to analyze pedestrian safety using video and other sensor streams. “We’re actually able to detect not only when accidents occur, but potentially different types of behavior that might make people more prone to injury,” Lien said.
In industrial environments, customers run Newton near their physical assets. “One of the main differentiators of physical AI from other LLM applications is that customers want their inference to run where their data lies,” she said. Most deployments today run on-premises or at the edge. “At the minimum, most of them want the inference and our platform to run on-prem… and have data sovereignty over that.”
Archetype has also explored early research collaborations in healthcare. In one project with UCSF orthopedics, Newton analyzed human motion data for indicators associated with conditions such as Parkinson’s disease or osteoarthritis. “This is still very nascent… but pointing to potential future applications in that domain,” she said.
Scaling Up
Lien said the company’s priority is scaling its platform and broadening its range of agents. “We have early users, we have an initial base of customers, but we really want to scale so that it can address a wider customer base,” she said. The team is also expanding template agents to make them more configurable and is continuing to invest in model development.
Archetype’s broader goal is to deliver a single model that can operate across an enterprise’s full set of physical assets. “Enterprises have a large number of physical assets, and having one foundation model that can solve all of these is of great value to them,” Lien said. “If we can see rapid expansion within those enterprises across multiple use cases and across multiple assets, that will be a very good indicator that indeed the values that we hypothesize we can bring are coming to fruition.”
