AI is breaking free from the chat box. As physical AI increasingly becomes a part of our workplaces and the world around us, it’s driving an evolution in robotics. It’s no longer enough to simply bolt AI onto an existing mechanical form. Instead, the equation is flipping: the shape, movements, operation, and, most importantly, senses of modern robots are built around AI. 

The importance of sensing in the era of physical AI is driven by a simple fact: A robot that can’t feel will eventually fail. And feeling doesn’t take place in a large language model (LLM) or a GPU in a data center. It takes place in the real world, at the edge. 

No matter how powerful a GPU or advanced an AI model is, they can’t underpin effective robotics without trustworthy data on the world around them. In other words, physical AI needs real-world inputs — from temperature and pressure to voltage, weight, and acoustics. Providing that additional sensory layer means that there must be an evolution in how we design and build robots. 

Creating a New Architecture for Physical AI 

Physical AI calls for a fundamentally different architecture from the robots of yesterday, one split into multiple different layers. Multimodal sensors act as the nervous system and reflex layer. These give the robot awareness of everything, including force, sound, and temperature.

High-quality data capture is essential here, because such data is fed into the perception and AI layer as raw material. Raw inputs are transformed into meaningful information that can be used for features like object detection, pose estimation (calculating the position of the robot in space), and more.  

All of this sensory input is captured before the robot acts. Movement, based on guidance from the AI that’s grounded in trusted sensory data, is controlled by a separate layer. At this point, high-performance compute handles the planning and motion goals, which are then sent to real-time motor controllers that ensure every movement is precise. 

Finally, edge gateways and cloud services capture multi-sensor trends, using them to predict maintenance and monitor use — all of which can be used to optimize performance. In other words, in physical AI architecture, multimodal sensor fusion is vital at both the start and the end of the loop. 

Why Sensor Fusion Matters Now

Crucially, in effective physical AI architecture, sensors don’t work in isolation. Sensor fusion — the combining of data from multiple sensors — is used to create a clearer picture of the real world. It’s rapidly replacing legacy systems that rely on single sensor monitoring. 

Picture a large robotic arm working in the automotive industry. Deep within its joint, a tiny 2- to 3-mA current offset pairs with a minor thermal change in torque; a minuscule hint that the bearings are starting to wear. With the right sensor fusion that combines temperature and torque, those tiny warning signs can be caught early by an AI. As a result, an engineer repairs the arm, and it keeps operating without failure. 

The same logic applies beyond automotive arms. Moving from single sensors to multi-sensor fusion creates dramatic improvements in autonomy and reliability across different robotics use cases. By fusing temperature, current/voltage, position, and other sensory input, anomaly detection improves and unexpected emergencies are reduced. 

This only becomes more important as new forms of robotics, e.g., humanoids, enter the workforce. For humanoid platforms to operate effectively, internal measurements like those focused on temperature and torque are fused with environmental sensing of the world around them. 

The Importance of Precision  

Achieving reliable sensor fusion requires more than simply connecting multiple sensors to an AI processor. The analog signal chain must preserve information across a wide range of sensor outputs, from microvolt-level bridge sensors used in force and torque measurements to industrial voltage and current signals employed for system monitoring. 

High-performance analog front ends (AFEs) are essential here. They integrate precision excitation sources, programmable gain amplification, low-drift references, and high-resolution data conversion into a common measurement platform, ensuring consistent accuracy across different sensing modalities. 

Devices like NXP's NAFEB43388 do just this, enabling a single hardware platform to acquire force, temperature, current, voltage, and resistance-based measurements. Such an architecture allows subtle correlations between mechanical, thermal, and electrical parameters to be detected with confidence, helping AI models identify early indicators of wear, efficiency degradation, or abnormal operating conditions before they develop into system-level failures.

With the right hardware, fusion becomes practical at scale, leading to efficiency and safety gains across fleets of robots. That might mean assembly robots or humanoid robots using vibration and motor currents to detect bearing wear. Or in another use case, autonomous mobile robots (AMRs) could capture wheel motor currents, encoder drift (the accumulation of small errors over time), and thermal signatures to build a live health profile across a whole fleet.