Robot hands have been operating mostly blind to physical sensation, relying on vision and language to guide their movements. But a Hong Kong startup is changing that equation by building the largest dataset of tactile feedback ever created for robotic manipulation. DAIMON Robotics released Daimon-Infinity this April, a comprehensive dataset featuring over one million hours of multimodal data with ultra-high-resolution touch sensing, spanning tasks from folding laundry at home to assembly line work in factories.

The move signals a fundamental shift in how the robotics industry thinks about embodied AI. While most robot systems today rely on the Vision-Language-Action (VLA) model, DAIMON's team has pioneered an alternative called Vision-Tactile-Language-Action (VTLA), which treats touch as equally important as sight. The company has open-sourced 10,000 hours of this dataset to accelerate research across the broader community.

Why Can't Robots Just Use Their Eyes?

The answer lies in the nature of physical manipulation. Robots performing delicate tasks, like picking up an eggshell or assembling precision components, need feedback about contact forces, material properties, slip, and friction. Vision alone cannot provide this information. Without tactile sensing, robots struggle with tasks that require fine motor control or adaptation to unexpected resistance.

"Tactile information is essential for providing feedback about contact states so that robots can guide their hands and fingers to perform reliable manipulation. Without tactile sensing, robots cannot reliably perform dexterous manipulation tasks," explained Prof. Michael Yu Wang, co-founder and chief scientist at DAIMON Robotics.

Prof. Michael Yu Wang, Co-founder and Chief Scientist at DAIMON Robotics

Prof. Wang, who earned his PhD at Carnegie Mellon studying manipulation and later founded the Robotics Institute at Hong Kong University of Science and Technology, has spent roughly four decades in the field. His insight is straightforward: humans rely heavily on touch when manipulating objects, and robots need the same capability to operate effectively in unstructured, real-world environments.

How Does DAIMON's Tactile Sensor Technology Work?

DAIMON's breakthrough is in the hardware itself. The company developed a vision-based tactile sensor that packs over 110,000 effective sensing units into a fingertip-sized module. Rather than relying on force sensors alone, this approach captures deformation, slip, friction, material properties, and surface textures, enabling comprehensive reconstruction of physical interactions.

The dataset was built through a distributed, out-of-lab collection network rather than a centralized factory. This approach allows DAIMON to generate millions of hours of data annually from diverse real-world environments, making the dataset more representative of actual working conditions.

Steps to Accelerate Embodied AI Development With Tactile Data

• Integrate Multimodal Fusion: Combine tactile feedback with vision, motion trajectories, and natural language instructions to create training-ready datasets that reflect how robots actually interact with objects in complex environments.
• Build Distributed Data Collection Networks: Move beyond centralized data factories to gather information across diverse real-world scenarios, including homes, factories, and service environments, ensuring models learn from varied conditions.
• Open-Source High-Quality Datasets: Release substantial portions of proprietary data to the research community to accelerate the entire field's progress, creating a shared foundation for embodied AI development.

Who Is Using This Data, and What Are They Building?

DAIMON's dataset development involved collaboration with leading research institutions and technology companies worldwide. Partners include Google DeepMind, Northwestern University, the National University of Singapore, and China Mobile. These organizations are using the tactile-rich data to train models tailored to their specific use cases, from academic research to commercial robotics deployment.

The dataset spans over 80 real-world scenarios and captures more than 2,000 human skills, providing a rich foundation for training general-purpose robotic foundation models. This breadth is critical because robots need to learn not just individual tasks, but how to adapt and generalize across different objects, environments, and manipulation challenges.

"Data scarcity remains a primary bottleneck in robot learning, particularly the lack of physical interaction data, which is essential for robots to operate effectively in the real world. By building and open-sourcing the dataset, we aim to provide the high-quality fuel needed to power embodied AI," stated Prof. Michael Yu Wang.

Prof. Michael Yu Wang, Co-founder and Chief Scientist at DAIMON Robotics

What Real-World Applications Are Coming First?

Prof. Wang sees touch-enabled robots making their first real-world inroads in specific sectors. Hotels, convenience stores, and other service environments in China are early adopters. These settings require robots to handle objects of varying fragility, navigate cluttered spaces, and interact safely with humans. Tactile feedback is essential for all of these challenges.

Beyond consumer-facing applications, manufacturing and assembly remain critical use cases. Robots equipped with tactile sensing can adapt to variations in parts, detect assembly errors through force feedback, and perform quality control tasks that currently require human inspection.

Why Does This Matter Now?

The robotics industry has reached an inflection point. Hardware for embodied AI is becoming more capable and affordable, but the bottleneck has shifted to software and training data. Without high-quality datasets that include tactile information, robots cannot learn the nuanced skills required for real-world deployment. DAIMON's decision to release 10,000 hours of data publicly reflects a recognition that the entire field benefits when foundational datasets are shared.

This approach contrasts with earlier robotics development, where companies guarded proprietary datasets closely. The shift toward open-sourcing reflects both the maturity of the field and the recognition that embodied AI requires collaborative progress. As more robots enter homes, factories, and service environments, the quality of their training data will directly determine their reliability and safety.