Tactile sensing is becoming a foundational capability for the next generation of embodied intelligence. For two years the robotics industry poured resources into bodies, large models, dexterous hands and vision. As those capabilities advanced quickly, the bottleneck narrowed to the contact layer: a robot must not only locate an object, but feel force, friction, slip and deformation while grasping, plugging, assembling and handling flexible materials.
On 10 August, Daimon Robotics released its tactile-anchored world model, Daimon-TWM. One day later it closed a strategic financing round worth several hundred million yuan, its second nine-figure round in two months. The moves point to one shift: tactile sensing is moving from the fingertips across the whole body and into the model layer.
Other leaders are converging. PaXini launched the PX-FOOTRIX multidimension foot tactile sensor. In July, 1X added tactile skin to the fingertips and palm of its new NEO hand. Boston Dynamics’ Atlas and Figure’s 03 list tactile sensing as baseline configuration. For leading firms, the question is no longer whether to use touch, but whether touch can move from a hardware signal into the model and become a base variable for understanding the physical world.
From sensing contact to predicting consequence
Daimon’s notable move is wiring its visuo-tactile sensor, native tactile data, VTLA model and Daimon-TWM into a single pipeline, so the model learns to understand, predict and exploit touch. At WAIC it ran two tasks: packing fragile fruit while controlling force, and organising a pencil case, where the zipper resists unevenly. When this writer placed an extra marker on the table, the robot paused, re-judged the space and adjusted. That pause is the valuable part, because in real factories and homes, positions, hardness and friction are never fixed.
Consider the most common factory task, a USB plug insertion. Once the plug nears the port, the key contact is occluded; a camera cannot tell whether it is aligned, deflected or stuck. A human finishes the alignment by feel. A conventional VLA model tends to generate a motion block, then execute it, so a small slip or jam accumulates. Daimon-TWM puts touch into the main loop of understanding, prediction and control: it judges the contact state, simulates what happens next, then corrects at 100 Hz. If the plug jams, it withdraws; if the object starts to slip, the fingers change force at once.
Daimon’s UniTacVLA paper tested eight tasks, adjusting, wiping, inserting and assembling. With no disturbance, success averaged 64 per cent; with added disturbance, about 53 per cent. A comparison model scored 26 per cent and 6 per cent. Ablation showed that simply adding tactile input helps little; success climbs only when the system understands the contact state, predicts future touch and corrects at high frequency.
No common language yet, but a route is chosen
Robotics touch still has no unified technical route. Piezoresistive and capacitive suits large-area coverage; Hall and magnetic arrays measure multidimension force; visuo-tactile reads deformation through a camera. Daimon’s bet is visuo-tactile, because the images and deformation fields it produces are the data that existing AI knows best, letting years of computer-vision encoders and training frameworks transfer directly.
Daimon’s answer is a “unified tactile token”: turn each tactile modality into a token the model can read, and let touch run through perception, prediction and control instead of sitting as an isolated input. Rather than forcing different sensors to produce identical data, it aligns their judgement of an object’s softness, roughness and protrusion, so the model forms a consistent physical reading even when raw data differ.
A business before the science is settled
Daimon’s underrated strength is shipping while researching. Its visuo-tactile devices have reached ten-thousand-unit shipments, serving more than 200 customers worldwide, over 50 of them overseas, with hardware delivered to OpenAI, Figure, Physical Intelligence, Meta, BMW and Google DeepMind. With China Mobile it launched a “data collection into homes” network; the first embodied-data 5S store opened in Chenzhou, Hunan, with 1,000 devices in phase one and a target of 1 million hours of real operation data per year.
The flywheel is visible: sensors bring revenue and customers, data devices produce native tactile data, data trains the model, and the model returns to validate on 3C and automotive lines. If it closes, Daimon sells more than a sensor each time; every delivery feeds a new task and new data. But the flywheel stalls at any link if customers stop repurchasing, the data proves unusable, or proofs of concept never become production orders. Full-stack capability is not the same as a profitable layer at every step.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience.