Introduction
In a world where robots are increasingly capable of seeing and understanding language, one critical sense remains largely underdeveloped: touch. DAIMON Robotics, a Hong Kong-based company founded just two and a half years ago, is working to change that. This April, the company released Daimon-Infinity, described as the largest omni-modal robotic dataset for physical AI. Featuring high-resolution tactile sensing, the dataset spans over 80 real-world scenarios and encompasses more than 2,000 human skills, from folding laundry at home to precise tasks on factory assembly lines. The project is a collaborative effort involving partners such as Google DeepMind, Northwestern University, and the National University of Singapore.
The Challenge of Robot Touch
Current robotic manipulation relies heavily on the Vision-Language-Action (VLA) model, which prioritizes visual and linguistic inputs while largely ignoring tactile feedback. This leaves robots “insensitive” to the physical nuances of objects—gripping too tightly, slipping, or failing to adjust to different textures. DAIMON aims to address this gap by elevating tactile perception to a level on par with vision. Their approach is built on years of research and a proprietary vision-based tactile sensor that packs over 110,000 effective sensing units into a fingertip-sized module. This sensor delivers monochromatic, high-resolution data that can capture subtle pressure changes, surface textures, and contact dynamics.
The Daimon-Infinity Dataset
The release of Daimon-Infinity marks a strategic shift for the company. Instead of focusing solely on product development, DAIMON has chosen to open-source 10,000 hours of its collected data to accelerate research globally. The full dataset includes million-hour scale multimodal data with ultra-high-resolution tactile feedback, collected through a distributed out-of-lab network capable of generating millions of hours of data annually. This resource is designed to help the AI community build more robust manipulation models that can generalize across tasks and environments.
Key Features of the Dataset
- Omni-modal richness: Combines vision, language, action, and now touch into a single comprehensive resource.
- Diverse scenarios: Data from over 80 real-world settings, including homes, factories, and commercial spaces.
- Human skill capture: Over 2,000 distinct manipulation skills performed by humans, providing a basis for imitation learning.
- Open access: A significant portion is freely available to the research community.
The VTLA Architecture: Touch as a First-Class Modality
Central to DAIMON's strategy is the Vision-Tactile-Language-Action (VTLA) architecture, pioneered by the company's co-founder and chief scientist, Prof. Michael Yu Wang. VTLA treats tactile data as an equal modality alongside vision and language, enabling robots to integrate touch information directly into their decision-making processes. This architecture is a fundamental departure from VLA models, which treat touch as a secondary or absent channel. By incorporating high-resolution tactile feedback, VTLA allows robots to dynamically adjust grip force, detect slippage, and adapt to object variations in real time.
Expert Insights and Strategic Vision
Prof. Michael Yu Wang brings decades of expertise to this initiative. After earning his PhD at Carnegie Mellon under the guidance of manipulation expert Matt Mason, he founded the Robotics Institute at the Hong Kong University of Science and Technology. An IEEE Fellow and former Editor-in-Chief of IEEE Transactions on Automation Science and Engineering, Wang has spent roughly 40 years advancing robotic manipulation. In discussions with DAIMON, he emphasized that the missing “insensitivity” in robot manipulation can only be solved by making touch a primary sense. The VTLA architecture and the Daimon-Infinity dataset are the first steps toward a future where robots can handle delicate tasks with human-like dexterity.
Real-World Deployment and Impact
DAIMON sees immediate applications for touch-enabled robots in commercial environments across China, including hotels, convenience stores, and light manufacturing. The ability to sense texture and pressure will allow robots to fold laundry safely, handle fragile items, and assemble components with precision. As the dataset grows and models improve, these robots will move from controlled labs into unstructured, everyday settings. The open-sourcing of data is intended to spur collaboration and reduce the time to deployment for embodied AI systems worldwide.
Expected Use Cases
- Household tasks: Folding clothes, washing dishes, preparing food.
- Manufacturing: Assembly line tasks requiring delicate touch, such as inserting components or handling glass.
- Retail and hospitality: Stocking shelves, cleaning surfaces, interacting with customers.
Conclusion
DAIMON Robotics is not just building better grippers; it is redefining how robots perceive and interact with the physical world. By releasing the Daimon-Infinity dataset and championing the VTLA architecture, the company offers the research community a foundation to develop truly tactile-aware AI. The path from insensitivity to sensitivity is long, but with massive data and a clear vision, DAIMON is giving robot hands the sense of touch they have long lacked.