Physical AI 需要从传感器到芯片的神经形态路径

The physical world does not wait for a sampling clock. It changes, moves, touches, slips, and vibrates continuously, yet the information that matters is often sparse, unpredictable, and fleeting.

A robot needs to react when an object approaches. A gripper needs to know the instant contact is made or slip begins. Yet much of today’s sensing and AI stack still samples physical signals at fixed rates, turns them into regular data, and sends them to processors built for dense computation.

This mismatch becomes particularly acute in high-speed applications—from autonomous drones to industrial processes such as laser additive manufacturing—where relevant physical dynamics can evolve faster than conventional sensing and processing pipelines can handle.

Sampling fast enough to capture these dynamics produces massive data streams, much of which is irrelevant to the task, making real-time processing increasingly constrained.

Turning sensing into selection

The path from physical sensing to AI needs an intermediate layer—not to process more data, but to select the information worth moving downstream. Neuromorphic principles provide a natural foundation for such a layer.

Event-based sensors turn sensing into selection. They respond to change as it happens, encoding only events that carry new information. An event camera, for example, records changes in light at each pixel, capturing their timing and location without continuously sampling the scene.

The same principle extends to touch, pressure, vibration, acoustic, and other modalities, where meaningful signals are sparse and asynchronous.

Crucially, an event is rooted in a physical change at a specific location and time. There is no global sampling clock deciding when information becomes available; events emerge when the world changes. The representation therefore follows the dynamics of the physical process itself.

Transforming raw events into physical state

The irony is that converting events to representations compatible with conventional processing architectures can erode much of the efficiency gained at the sensor. Once timing, location, and other metadata are appended and serialized for downstream processing, a sparse physical event can expand into tens of bits of conventional digital data.

This is where the sensing-to-AI layer becomes important. It transforms physical events into task-relevant state. An individual event says little; meaning emerges across space and time. Individual events say little; meaning emerges from their spatial and temporal relationships—within a sensor and across modalities.

For a drone, the spatial and temporal pattern of visual events across pixels in an onboard event camera can reveal an approaching obstacle, relative motion or occlusion.

The layer converts that raw event stream into a compact physical state—for example, obstacle proximity, relative velocity, direction of motion, time to collision—so downstream AI receives the information it needs without carrying the full event stream and its associated metadata through the pipeline.

Lessons from nature

Nature follows a similar principle. The brain does not process the world as raw sensor data. It receives an interface to the world. A fly, an owl, a fish, and an elephant do not expose the same physical world to their brains. Their bodies, sensors, and early neural circuits transform different physical signals into representations suited to what each animal can do.

The goal is not to replace GPUs or NPUs, but to feed them better information. These processors remain highly effective at dense computation—including neural-network inference, prediction, planning, and reasoning.

The proposed layer complements them by filtering and structuring sensor data upstream, so GPUs and/or NPUs receive information suited to their computing strengths.

The layer could also trigger reflex-like responses with ultra-low latency—for example, initiating an evasive maneuver when it detects an imminent collision—while higher-level AI handles context, intent, and longer-horizon decisions. This creates a hierarchy of control: immediate physical response close to the sensor; context, intent, and planning upstream.

Sensor-to-logic co-design in silicon

Neuromorphic principles shift the boundary between sensing and processing. With sensor-to-logic integration, including 3D silicon stacking, events can increasingly be processed close to their physical origin and native timing, creating a hardware foundation for the proposed sensing-to-AI layer.

The deeper opportunity is co-design: The sensor determines what becomes an event; the sensing-to-AI layer determines what those events mean for the task. Together, they can align sensing and computation with the physical task.

Specialized silicon can perform event-to-task transformations directly at the sensor, eliminating unnecessary data movement, representation, and interface overheads.

This opens a new design space for semiconductor companies and design houses: optimizing not just how sensors sense, but what information they produce. Sensors, software stacks,

Physical AI gyms, and simulation frameworks provide the discovery environment. Real applications reveal which sensing-to-task transformations matter. The most valuable can become reusable IP and, ultimately, specialized silicon.

Standardizing the physical-state interface

Without a common structure, specialized sensing-to-AI hardware risks creating a new form of fragmentation. A common physical-state interface—analogous in spirit to an ISA—could provide a shared vocabulary across machines and modalities. It should not standardize sensors; it should standardize what AI needs to know—obstacle proximity, relative velocity, direction of motion, and time to collision.

One machine might infer an approaching obstacle through event-based vision, another through event-based radar, yet both could expose the same physical state to AI. Sensing stays specialized; the interface becomes common.

Driven by users, applications, and industry, the interface becomes the contract between embodied perception and AI, while event-native specialized silicon provides the efficient path from physical events to that interface.