Skip to the stories
Robotics & Automation

Physical AI Needs Touch, Edge Processing and Better Robot Training Data

  • Majeed Ahmad
  • 657 reads

Robots need tactile feedback and local processing to handle variable tasks; industrial applications may lead adoption before humanoids.

Physical AI Needs Touch, Edge Processing and Better Robot Training Data

Robots are beginning to take on work that conventional programming struggles with: handling objects whose shape, stiffness or position varies from one cycle to the next. Making that practical depends on more than better vision. Machines also need to sense contact, judge force and respond quickly where the action happens.

That was the focus of a discussion at Synaptics Tech Day 2026, moderated by EE Times executive editor Nitin Dahad. Participants included 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.

Touch supplies information vision misses

For a robot, recognizing a glass is only part of the job. It must also apply enough force to lift it without crushing it, and detect if it begins to slip. The same challenge appears when handling fruit or an egg: the right grip depends on the object’s physical properties and the changing contact between gripper and object.

Ganesan described physical AI as giving machines the ability to sense, perceive, adapt and act. Tactile sensors can turn contact into data that a processor can use to adjust a grasp. Duchaine said that vision alone doesn’t directly reveal friction, texture, stiffness or force. Touch can also make training data more useful by distinguishing a stable grasp from one where an object is slipping.

Mehta pointed to another limitation of camera-based perception: a robot may lose sight of its fingers when they’re behind an object, and vision can’t directly measure force feedback. TacnIQ is exploring three-dimensional force maps that describe deformation and identify soft and hard points. Such maps could give a robot information about an object even when the camera’s view is obstructed.

Training needs records of physical interaction

Robot training data often describes what an object looks like without adequately recording what happened during contact. Duchaine said that systems need tactile, force and torque measurements, along with better annotation of how objects relate to actions. Recognizing an object is not the same as knowing how to manipulate it.

Mehta outlined a three-stage approach at TacnIQ: collect tactile data, organize it into physical characteristics, then connect those characteristics to applications. The company has collected approximately 6,000 hours of data across five industries, representing more than 10 million contacts. Its data categories include slip, shear, contact duration and material properties. The aim is to learn underlying physical behavior so that a separate model needn’t be trained for every task.

Herson expects self-annotation to matter as data collection scales. Tactile sensors, inertial sensors and motor currents can be combined to generate records that check and reinforce one another. That could reduce dependence on people manually labeling every interaction and help improve dexterity.

Industrial work is the near-term proving ground

Chowdhury sees humanoid robots as a longer-term opportunity, with nearer-term uses in warehousing and manufacturing. Touch and force feedback are particularly relevant to deformable or irregular objects that are hard to handle with fixed, conventional programming. Existing system integrators can support those applications.

Ganesan noted that industrial customers handle products ranging from glass bottles to fruit and eggs, often with mechanical or vacuum-based grippers that work best when objects are uniform. He expects faster adoption of industrial applications around 2028 and 2029. His comparison is to autonomous vehicles: driver-assistance features reached the market before fully autonomous driving, and he expects individual robotic capabilities to spread before fully capable humanoids.

Packaging is already one area where Duchaine sees deployments taking shape, particularly where cycle time isn’t the main constraint. He expects the next step to be more dexterous robots that outperform conventional industrial systems on variable tasks, without needing a humanoid form. Better speed and reliability would extend the range of work they can take on.

Local responses, coordinated systems

Combining sensors is only useful if the robot can process their signals in time. Ganesan expects vision to work alongside other sensing modalities, with some computation close to the sensors—in a robotic hand, for example. Local processing can handle immediate responses while a central processor manages higher-level decisions, an arrangement he compared with the human nervous system.

Duchaine said that continuously processing large volumes of visual data can introduce delays and produce jerky movement. Touch and force sensors provide lower-dimensional signals that can help a robot react more smoothly. The practical architecture uses vision and physical feedback together, with each supplying different information.

Broader deployment will also depend on making robots easier to integrate and economically viable. Herson identified lights-out manufacturing, including a large warehouse operating without employees on site, as a significant milestone. He said it will require more standardized robotic functions and a wider deployment ecosystem. Chowdhury added that agentic workflows could help system integrators and smaller manufacturers coordinate robotic assets without requiring the extensive research resources of the largest companies.

Ganesan expects industrial settings to establish the foundation for wider use because robots can first perform tasks without navigating complex interactions with people. The practical signal to watch is whether tactile sensing, faster local responses and simpler integration make variable jobs repeatable at production scale.

Source: Majeed Ahmad · www.edn.com

From the same set