Daimon Robotics building a tactile-data pipeline is its answer to a practical problem: cameras cannot tell robots everything about physical contact. The company is connecting tactile sensors, data collection, simulation and artificial intelligence models so other developers can train robots for manipulation tasks.
“By building tactile intelligence, we’re actually playing the role as an infrastructure provider,” Partner and Vice President of Technology Dr. Zhu told The Robot Report. He said Daimon is not trying to become a foundation-model developer. Instead, it aims to supply tools and data infrastructure to companies building robotic systems.
Daimon Robotics building a pipeline for touch data
Useful tactile data can be difficult to collect at scale. Unlike camera footage, tactile information depends on physical contact and varies with an object’s material, geometry, applied force and task. Daimon is developing a pipeline to capture and process those interactions, with tactile feedback operating at 30 Hz.
The company is also linking real-world measurements with simulation. It calibrates simulations against physical measurements before transferring collected data into virtual environments. That work addresses differences between simulated and real settings, including how friction and contact forces behave.
“Let’s be frank; it’s not good enough. It’s kind of terrible,” Zhu said of the current state of simulation. Daimon is developing tactile simulations based on physical properties measured through real-world testing. The stated aim is to make simulations more faithful and make it easier to transfer learned capabilities to physical robots.
The company has discussed collecting more than 10,000 hours of data. Its longer-term Daimon-Infinity initiative targets more than 1 million hours of tactile embodied data. Zhu said collaboration with developers helps Daimon learn what they need from its sensors, data-collection devices, software and evaluation tools. “Doing everything ourselves was not enough,” he said.
From tactile demonstrations to industrial work
Daimon is targeting assembly and packaging, where robots may need to detect a slipping object, check whether a component is seated or adjust when a part behaves unexpectedly. Tactile feedback can help a system respond to contact rather than follow a fixed trajectory.
At IROS 2026, the company demonstrated delicate manipulation tasks, including applying heat-transfer designs to tote bags and threading bracelets. It also showed a robot collecting and packaging broken glass. Daimon said these demonstrations used foundation models trained on data from multiple scenarios, rather than models built only around individual robot demonstrations.
The company uses its Umi device to collect data from human interactions. It has identified Lenovo Capital and Inovance Technology as key partners. Zhu said Daimon has moved beyond proof-of-concept projects into industrial assembly lines and warehouses, but declined to name customers because of nondisclosure agreements.
Whether tactile sensing improves performance consistently in production remains a commercial test. Robots in those settings must handle differences in objects, materials and operating conditions. Related reporting on devices exchanging messages with patients and AI in clinical decision support describes other ways software and connected systems are being applied in healthcare and technology. Research coverage of a growing ALS data engine offers another example of infrastructure built around large research datasets.
Models still face a generalization test
Daimon’s Daimon-TWM effort focuses on a tactile-grounded world model intended to help robots reason about physical interactions. The company is also exploring whether data from human interactions can help train models for different robot configurations, including humanoids and dual-arm manipulators.
Developers still need to show that models work across different objects, materials, robots and operating conditions. Zhu cautioned that models trained on large datasets have limited ability to generalize. “The model is not as intelligent as we think,” he said. He described some generalization as in-domain, and said embodied-intelligence models remain in development.
He added that collecting data from similar tasks can help models generalize to new applications. Daimon’s next challenge is to show whether its connected sensing, data and simulation tools can support reliable performance across a broader range of real-world tasks.
