Soft robots that grip things underwater often sense touch with liquid metal: a gallium alloy sealed inside thin channels in silicone, whose electrical resistance rises when a contact load squeezes the channel. These sensors are calibrated where they are built, near atmospheric pressure. At depth, the surrounding water adds a uniform squeeze from every direction, called hydrostatic pressure, and two things change. The resting resistance shifts. More importantly, the silicone may respond differently to the same fingertip force, because many polymers stiffen or yield differently when compressed on all sides. A calibration model learned at the surface then extrapolates to a pressure it has never seen, and black-box models are known to fail at exactly this kind of extrapolation.
This note proposes to treat ambient depth as a measured, physically structured context variable instead of an unmodeled drift. A second, shielded liquid-metal channel that cannot be touched reads only the ambient pressure. The calibration model is then split into two parts. A flexible network learns the shape of the touch response from shallow data. Ambient pressure is allowed to change only an offset and a gain, and both must come from a small library of physically motivated functional forms. The claim to test is that this split extrapolates from shallow training pressures to pressures several times higher, while plain recurrent networks, dimensionless-number projection, and simple reference subtraction do not. The proposal also asks a prior question honestly: does the touch gain actually change with depth? If it does not, the simple subtraction trick is enough, and the paper should say so.
Core Claim
Hydrostatic loading and contact loading enter an elastomer differently. The mean stress, $p$ (minus one third of the trace of the stress tensor), compresses the material volumetrically. Silicone and liquid metal are both nearly incompressible, so in a sealed, fully filled channel $p$ alone should barely change the channel's cross-section. Contact forces produce shear-like (deviatoric) stress, and that is what pinches the channel. In a pressure-sensitive material, ambient pressure also shifts how the material yields and flows under that shear. Void growth and damage ahead of a crack in pressure-sensitive dilatant polymers shows this in finite-element simulations of glassy polymers. In that work, raising a pressure-sensitivity angle $\alpha$ from 0° to 20° doubled plastic dissipation at fixed load, and a mismatch between $\alpha$ and the dilatancy angle $\beta$ added roughly 60% more damage. Those are glassy polymers, not silicone, so they motivate the hypothesis but do not establish it for elastomers. Measuring whether the effect exists in a silicone sensor is part of the work.
The consequence for modeling is a specific structural prior. If depth mainly shifts and rescales the touch response, the sensor reading should be well described by
$$\Delta R(t) = a(p) + b(p)\, h_\theta(u_{\le t}),$$
where $\Delta R(t)$ is the resistance change of the touch channel and $p$ is the ambient pressure read from the shielded reference channel. The term $h_\theta$ is a recurrent network acting on the history of contact inputs $u_{\le t}$, which captures hysteresis and viscoelastic memory. The offset $a(p)$ and gain $b(p)$ are sparse combinations of a small library of terms, $\{1, p, p^2, \log(1+p/\mu)\}$, where $\mu$ is the shear modulus of the elastomer. The approach succeeds when this separable form holds and the library contains the right term. It should fail in a predictable way when contact loads become large enough to cause pressure-dependent yielding, because the response then stops being separable. Locating that breakdown is a result in its own right.
Relation to Prior Work
The sensing hardware comes from 3D Printable Soft Liquid Metal Sensors for Delicate Manipulation Tasks. That work fills channels in printed Silicone 40A with Galinstan, a gallium-indium-tin alloy, and reads resistance with 4-wire sensing. It reports contact forces below 0.5 N detected underwater and stable output over 500 grasps. Its tests were shallow, it used no learned calibration, and it had no notion of depth as a shift in operating conditions. All three are what this proposal adds.
The only learned soft-sensor calibration in the file is Soft Robots' Dynamic Posture Perception Using Kirigami-Inspired Flexible Sensors with Porous Structures and Long Short-Term Memory (LSTM) Neural Networks. It fits an LSTM to conductive-sponge bending sensors and reaches 3.85° RMSE with $R^2 = 0.97$, but only on a random 70/30 split of a single recording. That number shows what a recurrent network achieves when test conditions match training. It says nothing about extrapolation.
The modeling choice comes from the out-of-distribution surrogate literature. Late Fusion Neural Operators for Parameterized Partial Differential Equations shows that feeding physical parameters into a network alongside the state entangles the two. It instead fuses parameters late through a sparse function library, which cut normalized RMSE on 1D advection by 32–86% relative to the next-best method. We borrow that late-fusion-plus-library structure directly. Ours differs in three ways: the parameter is measured online by a reference channel rather than given, the backbone is a recurrent model of a hysteretic material rather than a PDE operator, and the data are physical measurements. Disentangled Generative Models for Robust Prediction of System Dynamics supports the same diagnosis from another angle: its supervised latent separation of physical parameters reduced error by over 30% where standard models drifted.
Two papers constrain the design. Enforcing Analytic Constraints in Neural Networks Emulating Physical Systems shows that obeying physical laws is necessary but not sufficient for extrapolation. That is why our prior is a restriction on how $p$ may enter the model, not a conservation penalty. Buckingham π-Invariant Test‑Time Projection for Robust PDE Surrogate Modeling maps test inputs onto equivalent training inputs using dimensionless groups, and cut out-of-distribution error on 2D thermal conduction from 9.88 to 1.25 MAE with an FNO (a Fourier neural operator, a network that learns PDE solutions in frequency space). That trick works when test conditions are training conditions in different units. Here the dimensionless group $p/\mu$ itself moves outside the training range, so no equivalent training input exists. We test that prediction rather than assume it.
Out-of-distribution transfer of PDE foundation models to material dynamics under extreme loading is a warning about what large pretrained models deliver here. Transfer gains to solid mechanics peaked near 2×, and errors stayed concentrated at interfaces.
Research Program
The work has three parts.
Simulated dataset. We build an ABAQUS model of a single liquid-filled channel in an elastomer block. It uses a hyperelastic matrix with an optional pressure-sensitive yield surface parameterized by $\alpha$ and $\beta$, following the parameterization of Chew et al. Each case applies an ambient pressure up to 1 MPa, then an indenter load history, and records the channel cross-section and the implied resistance. Sweeping $\alpha$ from zero (pressure-insensitive) upward gives a controlled ground truth. When $\alpha = 0$, the separable model should be nearly exact. As $\alpha$ grows, it should degrade. This tells us in advance which regime the hardware occupies.
Physical dataset. We print touch-plus-reference sensor pairs in Silicone 40A following the Howard et al. process. The pairs are tested in a water-filled pressure vessel from atmospheric to about 1 MPa gauge, roughly 100 m of water. A motorized indenter inside the vessel applies load histories at three rates, and a load cell beneath it records the true force. Training data cover 0–100 kPa gauge. Test data cover 300 kPa to 1 MPa, held out entirely.
A secondary hardware variant borrows a stress-relief idea from rock mechanics. Research on the damage characteristics of rock masses based on double guide-hole blasting under high in-situ stress and Proactive intervention depressurization control for pre-mining roof fall and mining impacts use empty holes or grooves to redirect high stress away from a protected opening. Our variant adds sealed, empty relief cavities beside the channel. These give a second shift, in geometry, and let us ask whether a physical design change shrinks the gain drift that the model would otherwise have to correct. The same polymer study warns that voids also seed damage, so we track cyclic durability alongside sensing accuracy.
Model. The structured model above is fit with $\ell_1$ sparsity on the library coefficients. The recurrent network $h_\theta$ is shared across all training pressures. The functions $a$ and $b$ are identified from at least three training pressure levels, so that a pressure dependence can be fit at all.
Baselines
The naive reference is the calibration in current practice. A single surface-level fit of force against resistance is applied unchanged at depth, which corresponds to how the Howard et al. sensor would be used. Every method must beat it, and its error growth with depth measures how large the problem is.
The strongest competitor in practice is reference-channel subtraction. The shielded channel's resistance change is subtracted from the touch channel's, and the same surface calibration is then applied. This is the standard compensation used for temperature in strain gauges. It needs no learning and will remove any pure offset $a(p)$. It is the comparison that decides whether the learned gain matters. It gets the same sensors and the same data.
The strongest learned competitor is an LSTM given $p$ as an extra input, in the style of the Kirigami posture work. Its in-distribution figure there is 3.85° RMSE on bending angle, a different quantity, so it serves only as a sanity anchor. No reported result was found on Lacuna for any learned calibration of liquid-metal sensors under elevated hydrostatic pressure.
We also run Buckingham-π test-time projection wrapped around the same LSTM, using $p/\mu$ and a force-based dimensionless group. Its best reported figure is the 9.88 → 1.25 MAE thermal result, again in a different setting. A Late Fusion operator run in its original form, with $p$ supplied as the parameter, gives a fourth comparison.
All learned methods receive the same training pressures, the same input windows, and an equal hyperparameter search budget of 50 trials, scored on a validation split at the highest training pressure. None of them sees test pressures during tuning.
Two ablations isolate the key ingredient. The first removes the gain term, setting $b \equiv 1$, which leaves learned subtraction. The second replaces the sparse library with an unconstrained small network for $a$ and $b$, which tests whether the restricted library rather than late fusion alone does the work.
We count the proposal as winning if, at test pressures of 3× and 10× the training maximum, its force RMSE beats reference subtraction and the pressure-input LSTM by a margin larger than the spread across five sensor copies. We would also want its error to grow more slowly with depth than theirs.
Why not just…?
Why not just subtract the reference channel? This is a real threat. If depth only shifts the resting resistance, subtraction is sufficient, and our gain term is unnecessary complexity. The decisive measurement is the slope of resistance against indenter force at each ambient pressure. If that slope is flat across pressures within measurement noise, we report that subtraction suffices and the contribution narrows to the characterization itself. The pressure-sensitive-polymer results give reason to expect the slope to move, but they come from glassy polymers, so only the measurement can settle it.
Why not just train a big recurrent network on data at every depth? If one could collect dense data at every operating depth, that would work, and we include such an "oracle" model trained on test pressures as an upper bound. The point is that sensors are calibrated on a bench, and deep pressure testing is costly. Under a parameter shift, Late Fusion and SD-VAE both show networks that receive the parameter as an input degrade, and the pressure-input LSTM baseline measures how much that happens here.
Why not just nondimensionalize and project, as in Buckingham-π methods? That method helps when a test case is a rescaled copy of a training case. Deep water does not rescale a shallow press. It raises $p/\mu$ beyond anything seen in training, and the method's own authors note it assumes known, scale-invariant physics. We expect it to add little, and running it as a baseline will show whether that is true.
Why not just stiffen or thicken the encapsulation? A stiffer shell reduces the sensitivity to touch that the sensor exists for. It also adds resistance to the actuator's motion, a cost the Kirigami work quantified as low resistive torque being central to unobtrusive sensing. The relief-cavity variant is our cheaper physical alternative, and it is evaluated head to head.
Decisive Checks
The first check is physical, and it comes before any modeling. We measure the gain $b(p)$ across the vessel's pressure range on five sensor copies. That result decides whether the learned structure is needed.
The second check is in simulation. We plot extrapolation error against the pressure-sensitivity angle $\alpha$. The structured model should match an oracle near $\alpha = 0$ and break down as yielding couples pressure and shear. If it instead fails even at $\alpha = 0$, the separable form is wrong.
The third check is the held-out deep-pressure test on hardware, against the baselines and ablations above. We report force RMSE at 3× and 10× the training pressure maximum, plus a hysteresis-loop error under the fastest load rate.
As a methods sanity check, the late-fusion fitting code is first run on the 1D advection split from the Late Fusion paper. We confirm it lands within that paper's reported improvement range before trusting it on sensor data.
References
- Void growth and damage ahead of a crack in pressure-sensitive dilatant polymers
- 3D Printable Soft Liquid Metal Sensors for Delicate Manipulation Tasks
- Soft Robots' Dynamic Posture Perception Using Kirigami-Inspired Flexible Sensors with Porous Structures and Long Short-Term Memory (LSTM) Neural Networks
- Late Fusion Neural Operators for Parameterized Partial Differential Equations
- Disentangled Generative Models for Robust Prediction of System Dynamics
- Enforcing Analytic Constraints in Neural Networks Emulating Physical Systems
- Buckingham $\pi$-Invariant Test‑Time Projection for Robust PDE Surrogate Modeling
- Out-of-distribution transfer of PDE foundation models to material dynamics under extreme loading
- Research on the damage characteristics of rock masses based on double guide-hole blasting under high in-situ stress
- Proactive intervention depressurization control for pre-mining roof fall and mining impacts