Daimon Robotics builds tactile-data pipeline for robot makers

A camera can see a connector. It cannot tell a robot whether the connector has clicked home, or that a bag is starting to slip from its grip. That gap, between what machines see and what they feel on contact, is what Daimon Robotics has chosen to work on, and the company wants to sell the tools for closing it rather than the finished robot.

The company's approach is a pipeline. Its partner and vice-president of technology, Dr. Zhu, told The Robot Report that Daimon is “providing the infrastructure” for downstream customers, and is “not aiming to be a foundation model developer”. In practice, that means tactile sensors, data-collection devices such as its Umi device for capturing human interactions, simulation tools and AI models, all designed to work together. Tactile feedback, in its system, runs at 30 Hz, meaning 30 readings a second, which is what contact-heavy tasks need when a part behaves unexpectedly.

The bottleneck is data. Images can be captured easily with cameras; touch data depends on each physical contact and changes with an object's material, geometry, the forces applied and the task itself. Simulation is meant to multiply that data, but the gap between virtual and real physics remains wide, especially for friction and contact forces. Daimon's answer is to calibrate its simulations against real measurements before moving collected data into virtual environments. The company has discussed collecting more than 10,000 hours of data, and its longer-term Daimon-Infinity initiative targets more than 1 million hours of tactile embodied data. Those are goals and plans, not a dataset that has been delivered.

Daimon says it has moved beyond demonstrations. At the IROS 2026 conference, it showed robots applying heat-transfer designs to tote bags, threading bracelets, and collecting and packaging broken glass, tasks it says were driven by foundation models trained on data from several scenarios. It names Lenovo Capital and Inovance Technology as key partners, and says it is working in industrial assembly lines and warehouses, without naming a single customer. Independent verification of those deployments is not available, and the company itself frames the real test as performance in production.

What changes, and for whom? If the approach works, manufacturers building assembly and packaging robots would be able to buy tactile capability and training data instead of collecting it themselves, a step that matters most for smaller developers that cannot run their own data programmes. Daimon is also exploring how its human-interaction data could train models for humanoid and dual-arm robots. Whether that happens quickly depends on questions the sources leave open: how the tools perform outside the company's own demonstrations, and whether the 1-million-hour ambition is reached.