Physical AI: Where AI meets the physical world

Artificial intelligence (AI) is moving beyond the cloud and into machines that must perceive, understand, and interact with the physical world. Unlike AI systems that primarily process digital information, physical AI must enable machines to sense their surroundings, make decisions, and act in real time.

That shift introduces new engineering challenges. A robot may recognize an object through computer vision, but knowing how tightly to grip it requires additional sensing, physical-world training data, and processing that can respond to changing conditions.

At Synaptics Tech Day 2026, EE Times executive editor Nitin Dahad explored these challenges with Satish Ganesan, senior VP and GM 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 examined how physical AI is evolving, why tactile sensing matters, and what it will take to move intelligent machines from experimentation to practical deployment.

The following Q&A has been edited and condensed for clarity.

Where is physical AI creating value?

Nitin Dahad, EE Times: Where do you see physical AI creating value?

Satish Ganesan, Synaptics: Physical AI is fundamentally about enabling an inanimate object to sense, perceive, adapt, and act. The opportunity is to bring those capabilities into everyday applications, from industrial systems to humanoid robots, using multiple sensing modalities and processing information at the edge.

Andrew Herson, Nvidia: Manufacturing is one of the areas where I expect significant impact. It’s an environment with complex requirements, repetitive activities and large amounts of data, making it well suited for physical AI.

Vincent Duchaine, Robotiq: We’re starting to see a shift toward deployment. The opportunity is in automating tasks that were previously too complex or variable for conventional programming. Physical AI is beginning to make it possible to automate applications that require greater flexibility.

What does touch add that vision alone cannot provide?

Dahad: A key part of physical AI is touch sensing. What does touch add that vision alone can’t?

Ganesan: Think about gripping a glass, handling fruit, or picking up an egg. Each task requires a different amount of force. Touch sensing allows a robot to detect an object’s properties and apply the appropriate pressure. The goal is to translate physical interactions into information that processors can use to make decisions.

Duchaine: Vision has made tremendous progress, but advanced manipulation requires information that pixels alone cannot capture, including friction, texture, stiffness, and force. Touch also improves training data by showing not just whether a grasp succeeded, but whether an object was slipping or the grip was stable.

Aashish Mehta, TacnIQ: With vision alone, you can’t directly measure force feedback. Even determining finger position while holding a cup can be difficult because the hand and object obscure each other.

We’re exploring how to build three-dimensional force maps that capture deformation, soft and hard points, and other physical characteristics. That information could help robotic systems manipulate objects without relying entirely on vision.

What’s the biggest challenge in training robots for the physical world?

Dahad: What’s the hardest data problem in training robots to interact with the physical world? What data is needed to make dexterity reliable.

Duchaine: We’re not capturing enough information about physical interactions, including tactile sensing, force, and torque. We also need better annotation. It’s not enough to recognize objects. Models need to understand the relationships between objects and actions.

Mehta: We’re approaching this in three stages. First, we’re collecting large amounts of tactile data to characterize physical interactions, similar to how a baby learns through touch. We’ve collected approximately 6,000 hours of data across five industries, representing more than 10 million contacts.

Next, we’re organizing that information into physical characteristics such as slip, shear, contact duration, and material properties. Finally, we’re translating those characteristics into applications. The goal is to understand the underlying physics, so we don’t have to train a separate model for every task.

Herson: Self-annotating data will be critical. We can’t expect humans to manually annotate every interaction. By combining information from tactile sensors, inertial sensors and motor currents, robotic systems can generate data that reinforces and validates itself. That could help make dexterity more reliable.

Which applications are likely to benefit first?

Dahad: What will make robotic dexterity reliable enough for practical applications?

George Chowdhury, ABI Research: Humanoid robots represent a longer-term opportunity, but there are applications today in warehousing and manufacturing. Touch and force feedback are particularly useful for manipulating deformable or irregular objects that are difficult to handle through conventional programming. These are applications that existing system integrators can support.

Ganesan: Industrial customers need to handle everything from glass bottles to fruit and eggs, often using mechanical or vacuum-based grippers that require uniformity.

I see a parallel with autonomous vehicles. Driver-assistance features reached the market before fully autonomous driving. Similarly, individual robotic capabilities will be adopted before fully capable humanoid robots. I expect industrial applications to see more rapid adoption around 2028 and 2029.

How will multimodal sensing and edge processing improve robotic performance?

Dahad: How should sensing, perception, and processing be organized so robots can respond appropriately?

Ganesan: The future will involve multimodal sensing, combining vision with other sensors. To achieve low latency, some processing needs to happen close to the sensing elements, whether that’s in a robotic hand or elsewhere in the body.

It’s similar to the human nervous system, where nerve endings handle immediate sensing functions while more complex decisions happen elsewhere. We expect robotic architectures to evolve in a similar way, with local processing for immediate responses and central processing for higher-level decisions.

Duchaine: One challenge is the amount of visual data robots must continuously process, which can introduce delays and contribute to jerky movements. Touch and force sensing provide lower-dimensional signals that can simplify the problem.

It’s not about choosing vision over touch. The two complement each other, and physical feedback can help robots respond more quickly and smoothly.

What will signal that physical AI is ready for broader deployment?

Dahad: What’s one realistic deployment milestone you expect the industry to reach next, and what needs to happen to get there?

Herson: I’d point to lights-out manufacturing. A major milestone would be a large warehouse operating without employees on site. To get there, we need greater standardization of robotic functions and a broader ecosystem that makes deployment economically viable.

Duchaine: We’re beginning to see deployments in tasks such as packaging, where cycle time isn’t the primary requirement. The next step will be robots that are more dexterous than conventional industrial systems but don’t necessarily need to be humanoid. As they become faster and more reliable, they’ll address a wider range of applications.

Chowdhury: Broader adoption requires making robotics accessible beyond companies with enormous R&D resources. Agentic workflows could help system integrators and smaller manufacturers coordinate different robotic assets and reduce integration challenges.

Ganesan: Industrial environments will lead the adoption because robots can perform tasks independently without immediately needing to interact with people in complex settings. Those applications will establish the foundation for broader deployment and, eventually, more capable humanoid robots.

Related Content

The post Physical AI: Where AI meets the physical world appeared first on EDN.

The tchotchke