At IROS 2026, Daimon Robotics put Daimon-TWM through two autonomous manipulation demos: stringing a friendship bracelet and printing a canvas tote. Daimon called the event the model's first in-person public demonstration since its August launch.

What the two demos showed

For the bracelet task, the robot picked up engraved beads, aligned their openings and threaded them onto flexible string. Daimon said the work tested fine positioning, handling of deformable material and responses to changing contact.

For the tote task, the robot positioned a sticker, removed film, applied heat, allowed curing and put the item in its final place. The company described that sequence as a force-sensitive, multi-step operation. Daimon said tactile input helped the robot revise its actions as each task progressed. For bracelet stringing, the company contrasted that behavior with a fixed motion sequence.

How Daimon describes the model

Daimon characterizes Daimon-TWM as a world model grounded in touch. The company says it combines physical recognition, prediction and real-time control for dexterous manipulation.

Daimon co-founder Michael Yu Wang describes a Vision-Tactile-Language-Action architecture that gives touch a role alongside vision. The company uses the labels Physical Cognition, Predictive Decision-Making and Real-Time Control for the functions it says interaction data supports.

Daimon argues that vision does not provide all the information needed after contact. Its explanation points to deformation, friction, slippage and force as signals that a robot must track while acting.

The sensor and data layers

The hardware layer is Daimon's DM-Tac line of vision-based tactile sensors. Daimon reports more than 110,000 effective sensing units in a fingertip-sized module and perception across more than 12 tactile modalities. The company says its monochromatic sensor records deformation, slip, friction, material properties and surface texture.

Rather than emitting only force readings, the sensor captures image sequences showing changes at its surface. Daimon says those images can be used to infer forces and contact states in frameworks built around visual inputs. Wang said the design goal was to reproduce fingertip abilities such as identifying materials, sensing force distribution and tracking motion.

Data-Nexus sits between the sensors and world model. Daimon describes it as a workflow joining acquisition devices, networks, processing and evaluation so that interaction data can be prepared for model training.

The Daimon-Infinity dataset

Daimon released Daimon-Infinity in April and described it as the largest omni-modal robotics dataset for physical AI with high-resolution touch data. The company reports more than 80 real-world scenarios and over 2,000 human skills, and says it opened 10,000 hours of data to the public.

Daimon also says its distributed collection network can produce millions of hours of data annually. The company told IEEE Spectrum that academic institutions, research groups and leading humanoid robotics companies use its tactile devices.

Daimon sees general-purpose manipulation extending from household work to industrial assembly. Its examples include folding laundry and using tools to apply controlled forces to parts.

Sources

This article was researched and fact-checked against the following sources: