Skip to the stories
Robotics & Automation

Physical AI Needs Tactile Sensing and Edge Processing to Improve Robot Dexterity

  • Majeed Ahmad
  • 775 reads

Touch and force data can help robots handle variable objects, while local processing supports faster responses. Industrial tasks are likely to lead adoption.

Physical AI Needs Tactile Sensing and Edge Processing to Improve Robot Dexterity

Robots that can recognize an object still need to work out how much force to use, whether it is slipping and how to adjust as conditions change. That gap between perception and physical action is where tactile sensing, richer training data and edge processing could make physical AI more useful in industry.

At Synaptics Tech Day 2026, EE Times executive editor Nitin Dahad discussed those engineering problems with Satish Ganesan, Synaptics’ senior VP and GM of Edge Interface & Sensing; 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. Their discussion pointed to industrial settings as an early proving ground, with more adaptable robots and broader deployment depending on advances in sensing, data and integration.

Touch adds information that vision misses

Vision can locate an object and help a robot identify it. It cannot directly measure properties such as friction, stiffness, texture or contact force. Those details matter when picking up objects with different tolerances for pressure, from a glass to fruit or an egg.

Ganesan described touch sensing as a way to turn physical contact into information a processor can use to adjust a grasp. Duchaine added that tactile data can distinguish a stable hold from one where the object is slipping. That gives a training system a more useful record of a manipulation attempt than a simple success or failure label.

Mehta said vision also struggles when a robot’s fingers or the object itself obscure the contact area. His team is exploring three-dimensional force maps that describe deformation, soft and hard points, and other physical characteristics. Such maps could give robots another way to assess an object during handling, rather than relying on an unobstructed camera view.

Training needs records of physical interaction

Robot training data often lacks the force, torque and tactile information needed to make dexterity reliable. Duchaine said models also need to learn relationships between objects and actions, not just recognize objects. A system needs to connect what it senses at contact with what happens when it pushes, grips or moves.

Mehta outlined a three-stage approach: collect tactile data, organize it into physical characteristics, then use those characteristics in applications. TacnIQ has collected approximately 6,000 hours of data across five industries, representing more than 10 million contacts. The data is being organized around features such as slip, shear, contact duration and material properties.

The aim, Mehta said, is to capture underlying physical behavior so teams don’t need to train a separate model for every task. Herson argued that robots will also need to generate and check more of their own training data. Combining tactile and inertial sensors with motor-current measurements could let a system cross-check what happened during an interaction, reducing reliance on people to label every event by hand.

Industrial handling offers an early route to use

Warehousing and manufacturing already have tasks where robots must handle irregular or deformable objects. Chowdhury said touch and force feedback could help with these jobs, which can be difficult to solve through conventional programming and are within reach of existing system integrators. Packaging is one area where deployments are beginning, Duchaine said, particularly when cycle time is not the leading constraint.

Many industrial grippers are mechanical or vacuum-based and work best when objects are uniform. Ganesan pointed to the range of items industrial customers need to handle, including glass bottles, fruit and eggs. More responsive sensing could help automation cover a wider range of objects and tasks.

Ganesan expects industrial applications to see faster adoption around 2028 and 2029. He compared this progression to autonomous vehicles, where driver-assistance features reached the market before fully autonomous driving. In robotics, individual capabilities may gain acceptance before fully capable humanoid machines. Chowdhury likewise sees humanoids as a longer-term opportunity rather than the first place to look for practical results.

Local processing can cut response delays

Physical AI combines inputs from multiple sensors, and the architecture has to support timely reactions. Ganesan expects some processing to take place close to the sensors, such as in a robotic hand, while central processing handles higher-level decisions. The analogy is the human nervous system: immediate sensing and response can happen locally, with more complex decisions handled elsewhere.

Duchaine noted that continuously processing large volumes of visual data can introduce delays and contribute to jerky movement. Touch and force sensors produce signals that can simplify some immediate control decisions. In practice, vision and touch serve different roles: cameras provide a view of the surroundings, while contact sensing helps a robot adjust once it reaches an object.

Deployment depends on reliability and integration

Herson described a large warehouse operating without employees on site as a milestone for lights-out manufacturing. Reaching that point will require more standardized robotic functions and an ecosystem that makes deployment economically viable. Chowdhury said smaller manufacturers and system integrators also need accessible ways to coordinate different robotic assets; agentic workflows could help address integration challenges.

Duchaine expects robots with more dexterity than conventional industrial systems to take on additional tasks, even when they aren’t humanoid. Faster, more reliable operation would widen the set of jobs they can handle. Ganesan sees industrial environments as a foundation for this progression, since they let robots work independently before they need to manage complex interactions with people.

The near-term measure of progress, then, is not humanoid appearance. It is whether sensing, processing and integration can make flexible manipulation dependable in specific industrial jobs. Those deployments will show which tactile signals and training methods help robots cope with variation, and where physical AI can move beyond experimentation.

Source: Majeed Ahmad · www.edn.com

From the same set