PHANES AI, founded by Professor Yan Shao from the Harbin Institute of Technology Shenzhen, has released TouchWorld—a predictive and reactive tactile foundational model. This model grants robots the ability to sense touch, enabling them to perform complex physical tasks with high accuracy.
PHANES AI's work eliminates a critical gap in existing VLA (vision-language-action) models. These models can visually recognize whether a robot's finger is in contact with a button, but they cannot determine if that button has actually been pressed.
TouchWorld endows tactile probing with a dual function in robot control. Firstly, it has a predictive function: before executing any action, the model predicts not only the appearance of the scene after completing a subtask but also the corresponding tactile map. This map indicates which finger should feel pressure, where contact should occur, and with what intensity. Thus, the robot receives a physical target point that cannot be verified solely through vision.
Secondly, it has a reactive function: after establishing contact, the model continuously analyzes tactile signals and joint states. This allows for micro-corrections to adjust finger position, grip force, and wrist angle in real time, without requiring the high-level policy to re-plan.
When tested on real robots across six different tasks—watering plants, cleaning a table, inserting a fork, inserting a cup, scrubbing a pan, and retrieving napkins—TouchWorld demonstrated an average success rate of 65.0% under undisturbed conditions. With disturbances, including target movement and grasping interference, the success rate was 57.2%, exceeding the strongest baseline by 15.7 and 16.0 percentage points, respectively. Each task was trained on 200 teleoperation trajectories and evaluated during 100 tests on real robots.
The company asserts that tactile feedback cannot simply be added as another data modality to existing VLA models. Tactile data fundamentally differs from visual data in information density, timing, and processing speed. If fed into the same model, tactile signals might be suppressed by visual information. Instead, TouchWorld uses a separate architecture where tactile prediction and reaction operate parallel to the VLA pipeline, rather than within it. This allows the model to both anticipate the results of physical contact and react to unexpected interactions with the necessary speed for real-world manipulation.
