☄️ Tactile-Only World Model for Soft Robotic Manipulation in Fluid Environments

Existing world models learn physics by watching millions of hours of video. The recent wave of tactile world models and associated frameworks, such as VT-WM, OmniVTA, DreamTacVLA, and VTAM, has shown that adding touch to vision dramatically improves contact-rich manipulation. Yet every one of these systems treats touch as a supplement to vision. They use camera-based gel sensors on rigid grippers in air, where the tactile signal is a 2D image of a deforming membrane on a stiff fingertip.

In murky water or in bodily fluids, for example, vision often fails. Turbidity, lighting, and particulate scatter make cameras unreliable below a few meters. For soft robots operating in these environments, the primary channel for understanding the world isn't sight but touch.

No existing world model captures this to our knowledge as of June 2026. Fluid soft robotics breaks every assumption in the current tactile world model literature:

• the body itself deforms,
• the fluid couples with the contact signal,
• the tactile signal is a 1D strain distribution along a deforming rod and
• the interaction dynamics are governed by equations that pure data-driven approaches cannot learn from sparse real-world collection alone (according to our reasoning & assumptions at least).

What We Are Building

We are building the first purely tactile world model for soft robotic manipulation in fluid environments. Unlike existing visuo-tactile world models that rely on camera-based gel sensors mounted on rigid grippers, our approach operates without any visual sensing. Tactile perception arises from distributed strain measurements along a tendon-driven Cosserat rod manipulator, where the deforming body itself acts as the sensor.

A physics-informed neural network (PINN) embedding the coupled Cosserat rod equations and slender body hydrodynamics generates dense synthetic tactile fields decomposed into tendon actuation, hydrodynamic loading and contact contributions. A learned world model trained on this physics-grounded synthetic data, regularized by the governing conservation laws, predicts future tactile states from sparse real sensor readings and actuation inputs. This enables model-based control policies that plan through predicted touch in environments where vision is degraded or absent, such as underwater coral monitoring, where the robot must sense and respond to delicate contact through body deformation alone.

This architecture inverts the modality hierarchy of every existing tactile world model. VT-WM, OmniVTA, DreamTacVLA, and VTAM all assume vision as the primary modality with touch as supplement. We make touch primary and remove vision entirely. The rod's proprioceptive state (its shape, recoverable from strain readings plus rod kinematics) combined with the tactile contact signal is everything. This is closer to how real marine animals navigate: an octopus tentacle exploring in murky water operates through touch and proprioception alone.

The target platform is a tendon-driven Cosserat rod tentacle: a single manipulator arm on a jellyfish-inspired robot designed for coral reef monitoring and gentle interaction with fragile marine structures. The tentacle must navigate around branching coral, make contact without exceeding damage thresholds, and sense object properties through its own deformation. The tendon actuation provides clean separation between actuation forces and external forces in the tactile signal: the tendon contribution is fully known from the prescribed tension and routing geometry, the fluid contribution comes from slender body theory, and whatever remains must be contact.