Every act of understanding — by a human, an animal, or a machine — begins not in a reasoning engine but in the structured perception of physical reality. We argued this in 2019. The world is now calling it Physical AI.
Rodney Brooks at MIT argued in the 1980s that abstract symbol manipulation — "a brain in a jar" — could never produce genuine intelligence. He was right but incomplete. The question was not merely whether intelligence needs a body. It was: what kind of sensory data does that body need to produce so that intelligence can be genuine?
That question points to the sensor, not the processor. The quality of machine intelligence is bounded, first and foremost, by the causal fidelity of its perception.
The structure of human language — spatial prepositions, temporal relations, causal connectives — is not arbitrary. It mirrors the structure of physical experience. "Before," "behind," "causes," "follows" — these are not logical primitives. They are the sediment of millions of years of creatures navigating a physical world and needing to communicate about it.
A machine that learns language from text alone is learning the shadow. Physical AI means giving the machine the light source.
This is the argument that changes everything about how intelligent systems should be designed. If understanding is grounded in physical perception, then the sensor architecture is not a peripheral input device. It is a constitutive element of intelligence itself.
Which means that a sensor which delivers causally-impoverished data — a point cloud — is not merely a suboptimal component. It is an epistemic bottleneck on the machine's capacity to understand anything at all.
In July 2019, a patent application was filed that made a claim unusual for a technical document: that the design of a LiDAR system should be derived from first principles about how intelligence works, not from incremental improvements on existing hardware.
The specification of what became US 2026/0056318 A1 opens not with a description of optical components but with an epistemological argument. It states that intelligent machines — autonomous vehicles, robots, navigating systems — require remote sensing systems capable of delivering data from which genuine comprehension of the surrounding environment can be derived. Not detection. Not object labeling. Comprehension.
It then argues that current LiDAR architectures fail this requirement not because they are insufficiently precise, but because they are architecturally wrong. The point cloud — the industry's standard output format — discards the causal relationships between illumination events and detected returns at the moment of data production. The sensor knows which photon came from which direction at which time. The point cloud throws that away. The AI downstream must statistically reconstruct what the sensor already knew.
"The present disclosure provides a unified method of intelligently comprehending LiDAR data — applicable to all types of sensing approaches — so as to assist a decision-making module given the surrounding state conditions."
US 2026/0056318 A1 · ¶[0014] · Filed July 2019The participation matrix — the data structure at the center of this disclosure — is the answer to that architectural failure. It is not a post-processing enhancement. It is a reconception of what a sensor should output: not scatter, but causal structure. A record of which illuminating event caused which detected return, at a known time, in a known direction. The information the AI needs to reason about the world as a physicist would — not as a statistician must.
This argument was made in July 2019. Jensen Huang named it "Physical AI" at CES in January 2025 — five and a half years later. He described it as AI that can "perceive, reason, plan and act" in the physical world. He called it the next frontier. He was right about what it is. He was late to the argument about why.
We are not making a territorial claim. The concept that intelligence requires a body that collides with the real world goes back to Norbert Wiener's cybernetics in the 1940s, to Rodney Brooks' embodied cognition thesis in the 1980s, to the cognitive science tradition that language and thought are grounded in sensorimotor experience. What is new here is the specific, patented, architecturally-realized answer to the question those traditions raised: what must the sensor produce so that the body can think?
That answer — the participation matrix, the waveguide-aperture illuminator, the CFSRP modality, the single omnidirectional detector — is the technology at the center of physical-ai.com. Not a philosophical position. An engineered system, with a 2019 priority date, built to deliver what Physical AI actually requires at its sensory foundation.
The thesis that intelligence is grounded in physical reality has been eluded before. It has a rich lineage — from cybernetics to embodied cognition to ecological psychology. What is new is the translation of that thesis into a concrete sensing architecture, filed as prior art, with a 2019 priority.
The timeline below maps the intellectual progression from philosophical claim to commercial vocabulary — and marks where this work sits in that progression.
The 2019 patent application does not use the phrase "Physical AI" — the term did not exist as a mainstream label. What it does argue, explicitly, in paragraphs [0005]–[0014] and [0028]–[0043], is that intelligent machines require sensing architectures that produce causally-structured representations of the physical world, and that the participation matrix framework enables AI systems to reason about the environment with a rigor that point-cloud processing cannot match. That is the Physical AI thesis — stated five years before Jensen Huang named it at CES.
This is not a claim to have invented a philosophical idea. It is a statement of documented intellectual and engineering priority: the specific argument that the sensor architecture determines the ceiling on machine intelligence, and the specific engineered answer to that argument, was filed, dated, and published. The vocabulary arrived late. The work did not.
A chess engine computes. A surgeon understands. The difference is not speed or accuracy — it is that the surgeon's knowledge is grounded in physical reality: the texture of tissue, the resistance of bone, the spatial geometry of the body. Intelligence, in any meaningful sense, is not a property of symbol manipulation. It is a property of systems that are coupled, through their senses, to the structure of the physical world.
The richness of human language — its prepositions, its tenses, its causal structure — did not arise from abstract logic. It arose from creatures that moved through space, manipulated objects, tracked the movement of other creatures, and needed to communicate about these physical events. Language learned from text alone learns the map. Physical AI means giving the machine the territory.
In current machine learning, the sensor is treated as a data input device — as peripheral as a keyboard. This is wrong. The quality of sensory data determines not just the accuracy of the machine's outputs but the nature of what it can understand at all. A machine fed point clouds can classify objects. A machine fed causally-structured participation matrices can reason about events. These are not the same cognitive capacity. The sensor determines the ceiling.
Transformer architectures, large language models, and reinforcement learning systems have reached extraordinary sophistication. They are not the limiting factor in autonomous systems. The limiting factor is what they are fed. Point clouds, camera frames, and radar returns are impoverished representations of a rich physical reality. The next order-of-magnitude improvement in machine intelligence will come not from bigger models but from better-structured sensory data. We are building that data infrastructure.
When Jensen Huang described Physical AI at CES 2025, he was describing a market segment. We mean something more fundamental: any AI system that claims genuine understanding of the world must be grounded in physical reality. This is not a design choice. It is an epistemological requirement. The implication is that physical-ai.com is not building for a market niche. It is building the foundational sensory infrastructure that every intelligent machine will require — whether it drives, walks, flies, operates, or thinks.
L3/L4 autonomy requires not object detection but scene comprehension — understanding intent, predicting motion, reasoning about occlusion. Participation matrix data makes this possible. Point clouds make it a statistical approximation.
A robot that picks up an object and places it precisely is doing physics. It needs to understand the causal geometry of the scene — not classify it. CFSRP sensing gives robotics the physical grounding that language models alone cannot provide.
Traffic systems, industrial automation, precision agriculture — every application that requires a machine to act reliably in an unstructured physical environment needs a sensing substrate that preserves causal structure. That is what we build.
Surgical robotics, diagnostic imaging, patient monitoring — the physical world of the human body demands the highest-fidelity sensory data. The participation matrix framework applies to any domain where causal scene understanding is required.