Bridging the Gap: How Physical AI and Tactile Sensing are Bringing Intelligence to the Real World

Artificial intelligence is rapidly breaking free from the confines of digital servers and the cloud, moving directly into physical machines that must perceive, understand, and interact dynamically with the world around them. Unlike traditional AI systems that primarily ingest, process, and output digital data, physical AI requires hardware and software to operate in real time. These systems must continuously sense their physical surroundings, make complex decisions on the fly, and execute physical actions safely and effectively.

This major technological evolution introduces an entirely new tier of engineering challenges. While a modern robotic system might effortlessly recognize an object across a room using advanced computer vision, knowing how to interact with that object—such as calculating precisely how tightly to grip a delicate item without crushing it—demands entirely different capabilities. It requires multi-modal sensing, specialized physical-world training datasets, and high-performance edge processing capable of responding instantaneously to shifting environmental conditions.

To examine these breakthroughs and hurdles, EE Times executive editor Nitin Dahad recently sat down with industry leaders at Synaptics Tech Day 2026. The panel featured Satish Ganesan, senior vice president and general manager of Edge Interface & Sensing at Synaptics; Aashish Mehta, CEO of TacnIQ; George Chowdhury, principal analyst for robotics at ABI Research; Vincent Duchaine, co-founder and CTO of Robotiq; and Andrew Herson, head of ecosystem development for physical AI at Nvidia.

The discussion provided a comprehensive look at how physical AI is maturing, why touch is becoming just as critical as sight, and what industrial sectors must overcome to transition intelligent machines from experimental labs into practical, widespread commercial deployment.

Where Physical AI is Creating Immediate Value

The conversation began by exploring the sectors where physical AI is already beginning to prove its economic and operational worth. According to the panel, the overarching goal of physical AI is to transform inanimate objects and standard machinery into responsive systems capable of sensing, perceiving, adapting, and acting independently.

Satish Ganesan emphasized that the ultimate opportunity lies in integrating these advanced capabilities into everyday applications, ranging from automated industrial equipment to sophisticated humanoid robots, by leveraging multiple sensing modalities and executing heavy data processing locally at the edge.

Manufacturing, in particular, is emerging as a primary battleground for these technologies. Andrew Herson of Nvidia noted that manufacturing environments feature uniquely complex requirements, repetitive yet strenuous activities, and massive volumes of data, making them a natural fit for early physical AI implementations.

Meanwhile, Vincent Duchaine of Robotiq pointed out a broader operational shift toward active commercial deployment. He explained that the primary industry opportunity involves automating tasks that were historically far too complex or variable for conventional programming, allowing automation to expand into applications requiring unprecedented operational flexibility.

The Vital Role of Touch in Robotic Manipulation

While computer vision has dominated robotics headlines for years, the panel agreed that true physical interaction demands a sense of touch that pixels alone can never replicate.

Ganesan illustrated the limitation by comparing everyday human actions to machine capabilities. Picking up a fragile glass, handling a ripe piece of fruit, or lifting an egg each requires a dramatically different application of force. Touch sensing allows a robot to dynamically evaluate an object’s physical properties and apply the exact amount of pressure required. The ultimate objective is to translate physical interactions into raw data that onboard processors can immediately use for real-time decision-making.

Duchaine reinforced this perspective, noting that while vision technology has made staggering strides, advanced manipulation requires variables that simple visual feeds cannot capture, including friction, surface texture, material stiffness, and precise force vectors. Furthermore, tactile feedback significantly improves machine-learning training data by indicating not just whether a grasp initially succeeded, but whether an object is actively slipping or if the grip remains stable over time.

Aashish Mehta of TacnIQ added that vision alone cannot measure direct force feedback. Even basic tasks like determining the precise finger position of a robotic hand while holding a standard cup can become obscured because the hand and the object frequently block each other from camera views.

To solve this, Mehta’s team is actively developing three-dimensional force maps designed to capture physical deformation, distinct hard and soft contact points, and other

tactile characteristics. This rich stream of information allows robotic systems to manipulate objects effectively without relying exclusively on visual guidance.

Overcoming the Toughest Data Challenges

Training robots to operate reliably in the physical world presents some of the most difficult data hurdles in modern engineering. Duchaine pointed out that the industry is still failing to capture enough comprehensive data regarding physical interactions, including fine-grained tactile feedback, torque, and force. Furthermore, the robotics sector desperately lacks high-quality data annotation. It is no longer enough for an AI model to simply recognize an object; the underlying models must deeply comprehend the physical relationships connecting objects and actions.

Detailing how TacnIQ approaches this obstacle, Mehta described a three-stage strategy. The company first collects massive volumes of tactile data to characterize physical interactions, likening the process to the way a human infant learns through physical touch. To date, TacnIQ has gathered approximately 6,000 hours of specialized data across five distinct industries, representing more than 10 million distinct physical contacts.

In the second stage, this raw information is organized into specific physical characteristics, such as slippage, shear forces, contact duration, and material properties. Finally, these characteristics are translated into practical applications, ensuring that systems understand fundamental physics rather than requiring engineers to train entirely separate models for every single new task.

Nvidia’s Andrew Herson emphasized that self-annotating data will be the linchpin for future success. Because humans cannot possibly manually label every physical interaction, robotic systems must combine data streams from tactile sensors, inertial measurement units, and motor currents to generate self-reinforcing, self-validating datasets that drastically improve operational dexterity.

Practical Applications and the Path to Deployment

When turning to the timeline for practical deployments, George Chowdhury of ABI Research noted that while humanoid robots represent an exciting long-term vision, immediate commercial opportunities lie firmly within warehousing and manufacturing. Touch and force feedback are particularly invaluable for handling deformable or irregular objects that have traditionally resisted conventional automation techniques.

Ganesan drew a compelling parallel to the automotive industry’s path toward autonomy. Just as driver-assistance features successfully reached the consumer market long before fully autonomous self-driving cars became commonplace, individual robotic capabilities will be adopted incrementally well before general-purpose humanoid robots saturate the market. He projected that industrial applications will experience accelerated adoption phases around 2028 and 2029.

To achieve the low latency required for these systems, Ganesan argued that robotic architectures must mimic biological systems. By adopting a distributed nervous system approach—where local sensing and processing occur directly within nerve endings like robotic hands, while higher-level decisions are handled centrally—machines can respond instantaneously to physical disruptions. Duchaine echoed this sentiment, explaining that lower-dimensional tactile and force signals can simplify the immense data processing burdens associated with continuous visual feeds, allowing robots to react faster and move more smoothly.

Discussing the ultimate milestones that will signal physical AI’s readiness for broader commercial integration, Herson pointed toward lights-out manufacturing facilities operating entirely without on-site human workers. Reaching this milestone will require deep standardization of robotic functions and a mature ecosystem that makes large-scale deployment economically viable.

Chowdhury concluded that broader adoption ultimately depends on making advanced robotics accessible to companies outside of massive, well-funded research and development institutions. Through the use of emerging agentic workflows, system integrators and smaller manufacturers will soon be able to coordinate diverse robotic assets seamlessly, bringing the promise of physical AI from experimental testing grounds directly to the global industrial floor.