Robots have been learning to see for decades, but they're only now learning to feel. Three major developments announced this week reveal a fundamental shift in how the robotics industry thinks about embodied AI: touch sensors and tactile feedback are becoming as essential as cameras for robots that need to manipulate objects in the real world (Source 1, 2, 3).

Why Can't Vision Alone Solve Robot Manipulation?

For years, roboticists assumed that if a robot could see an object clearly, it could grasp and manipulate it. But real-world tasks like assembly, insertion, and delicate object handling require something vision cannot provide: feedback about physical contact. When a human picks up an egg, they don't rely on sight alone; they feel the pressure, detect slip, and sense the shell's fragility through their fingertips.

Daimon Robotics and Galbot, two companies focused on embodied AI, launched RobOmni at the IEEE International Conference on Robotics and Automation (ICRA) 2026 this week. The platform is the first standardized benchmark designed specifically to measure how much tactile sensing actually improves robot performance in contact-rich tasks.

"While vision enables robots to perceive the world, it cannot fully capture the physical interactions that underpin real-world manipulation, limiting robots' ability to operate reliably in unstructured environments," the companies explained in their announcement.

Daimon Robotics and Galbot, ICRA 2026 announcement

The gap between what robots can see and what they can do has been one of embodied AI's most stubborn problems. RobOmni addresses this by providing a reproducible framework that measures robot performance across multiple dimensions, not just task completion.

What Does RobOmni Actually Measure?

The benchmark evaluates robot performance across several key metrics that matter in real manufacturing environments:

• Task Success Rate: Whether the robot completed the manipulation task without failure.
• Manipulation Efficiency: How quickly and smoothly the robot performed the action.
• Dexterous Capability: The robot's ability to perform complex, fine-grained movements.
• Failure Events: Specific problems like slip, jamming, collision, or the need to retry the task.
• Generalization Robustness: Whether the robot could adapt to variations in object properties or task conditions.

The platform includes dozens of manipulation scenarios such as grasping, placement, precision insertion, and component assembly. These tasks were selected because they closely reflect operational challenges commonly encountered in manufacturing and service robotics.

A critical feature is tactile ablation testing, which allows developers to run the same task with and without tactile information. This enables researchers to directly measure how much tactile sensing actually contributes to task performance, moving beyond assumptions to hard data.

How Are Companies Integrating Tactile AI Into Real Robots?

While RobOmni provides the testing framework, other companies are building the hardware and software to make tactile sensing practical. TARS, an embodied AI company, unveiled its DexHand platform at ICRA 2026 this week, showcasing a hand designed from first principles to replicate human tactile capability.

DexHand features a 21-degree-of-freedom (DoF) architecture modeled directly on human hand anatomy. Unlike conventional robotic hands that use parallel-joint designs, DexHand replicates the spatial convergence of the thumb's joints, eliminating motion blind spots that typically limit dexterity.

The hand's fingertips integrate ultra-high-resolution miniature cameras capable of capturing microscopic textures as fine as 0.05 millimeters at over 240 frames per second. TARS' AWE 3.0 embodied foundation model enables the robot to understand physical properties such as hardness, roughness, and slip risk, and to predict problems rather than merely reacting after contact is made.

"TARS' DexHand is the optimized interface between human intelligence and robotic action," said Dr. Ding, TARS' chief scientist and co-founder.

Dr. Ding, Chief Scientist and Co-Founder at TARS

On the manufacturing side, DexHand's design uses just three motor types and reducer types, making it purpose-built for automated assembly lines. This focus on manufacturability suggests that tactile-enabled robots may finally be moving from research labs into actual factories.

What Strategic Partnerships Are Accelerating Tactile AI Development?

Beyond individual product launches, the industry is consolidating around partnerships that combine AI software with robotics hardware. VinDynamics, a robotics company within Vietnam's Vingroup conglomerate, announced a strategic partnership with Skild AI, a foundation model company backed by SoftBank Group, NVIDIA Ventures, and Jeff Bezos.

The collaboration will focus on embodied AI research, robot manipulation, sim-to-real transfer (the challenge of moving trained models from simulation to physical robots), edge AI deployment, and validation of humanoid systems in real-world environments. Both companies will explore integrating Skild Brain, Skild AI's omnibodied AI software, into VinDynamics' humanoid robots.

VinDynamics brings manufacturing scale and access to real-world deployment environments through the Vingroup ecosystem, while Skild AI contributes foundation model technology. The partnership aims to accelerate deployment of humanoid robots capable of operating in complex and dynamic environments across manufacturing, logistics, hospitality, and commercial services.

"The future of robotics depends on combining scalable physical systems with adaptable, general-purpose intelligence," said Deepak Pathak, co-founder and CEO of Skild AI.

Deepak Pathak, Co-Founder and CEO at Skild AI

How to Evaluate Tactile Sensing in Your Robotics Development

If you're involved in robotics development or manufacturing automation, understanding tactile sensing's role in your specific use case is increasingly important. Here are practical steps to assess whether tactile perception matters for your application:

• Identify Contact-Rich Tasks: Determine whether your target application involves grasping, insertion, assembly, or handling of delicate objects where physical feedback is critical to success.
• Measure Current Failure Modes: Track whether failures are primarily due to slip, jamming, collision, or inability to detect object properties; these are areas where tactile sensing provides the most value.
• Test with Standardized Benchmarks: Use platforms like RobOmni to evaluate how much tactile information would improve your specific manipulation tasks before investing in hardware integration.
• Consider Manufacturing Feasibility: Assess whether tactile sensors can be integrated into your robot's design without excessive complexity; TARS' approach using standardized motor and reducer types suggests manufacturability is becoming a priority.

What Does This Mean for the Future of Physical AI?

The convergence of three developments this week signals a maturation in embodied AI. First, standardized benchmarks like RobOmni are replacing fragmented, company-specific testing. Second, hardware designs like DexHand are moving from biomimetic novelty to manufacturing-ready products. Third, strategic partnerships are combining AI software with robotics hardware at scale (Source 1, 2, 3).

RobOmni will soon support real-robot validation, streamlining the sim-to-real pipeline. This represents the industry's most comprehensive standardized evaluation framework centered on tactile perception and dexterous manipulation, filling a critical infrastructure gap that has slowed progress for years.

The robotics industry has long established benchmarks for perception, navigation, and machine learning performance. Tactile intelligence, however, lacked a comparable framework until now. As physical AI continues to evolve, evaluation infrastructure is becoming just as important as advancing the underlying AI models themselves.