Passive Recurrent Cells for Real-Time Viscoelastic Tissue Forces

Surgical phantoms and fast learned tissue models share a hidden defect: they have no memory. Real soft tissue is viscoelastic. When you press on it and hold still, the reaction force decays over seconds. When you load and unload it, the force follows a different path on the way back. When you move faster, the tissue pushes back harder. Yet the learned models that run fast enough for haptic feedback (a robot pushing a felt force back into the surgeon's hand) map the current tool displacement directly to a force, as if the tissue had no history. Generic recurrent networks do have memory. A recurrent network is a model that carries a hidden state from one time step to the next, and gated variants such as the LSTM and GRU are standard. But nothing stops such a network from inventing energy. In a closed haptic loop, a force model that can output more energy than it absorbed can make the device oscillate or run away.

We propose a recurrent cell whose hidden state is, by construction, the internal stress of a bank of spring-and-damper branches with learned, deformation-dependent stiffness and relaxation times. The network keeps the flexibility of a learned model: nonlinear elasticity, time constants that depend on strain, and inputs from kinematics or vision. Its structure prevents it from generating net mechanical energy over a loading cycle, so it can be dropped into a 1 kHz haptic loop without destabilizing it. The claim to test is that this structure costs little or no accuracy against an unconstrained GRU on in-distribution data. It should beat the GRU when extrapolating to loading rates it has not seen, and it should never violate energy dissipation. A classical fitted viscoelastic model should not be able to match it once the tissue is nonlinear.

What exists and where the gap is

Two learned tissue models in the literature show the speed that is achievable and also the missing ingredient. Conditional Graph Neural Network for Predicting Soft Tissue Deformation and Forces predicts surface displacements and reaction forces under indentation in about 3 ms, with 0.37 ± 0.05 N force error. It does this by pretraining on 16,500 mass-spring simulations and fine-tuning on 244 phantom states. NNWarp adds a learned nonlinear correction on top of linear finite-element simulation and runs about 1,000× faster than nonlinear solvers. Both treat every state as a memoryless snapshot, and that is the property we change.

Recurrent models appear at the edges of surgical mechanics, and they motivate this proposal because each finds that memory matters. A recurrent convolutional neural network approach for sensorless force estimation in robotic surgery feeds CNN video features and tool kinematics to an LSTM and beats an ARMAX baseline on silicone phantoms. It works on a static, uniform phantom and does worse on pulling than on pushing. Roughly Collected Dataset for Contact Force Sensing Catheter shows that a GRU tracks "shift of reference," the failure of a fiber-optic catheter sensor to return to baseline after unloading because its structure relaxes. On that task a static network predicts phantom contact. Efficiently Calibrating Cable-Driven Surgical Robots uses an LSTM over a window of 4 steps to remove dVRK cable hysteresis, cutting tracking error from 2.96 mm to 0.65 mm. RUGNN puts GRUs inside graph message passing to stop error drift in multi-step sheet-metal forming. None of these constrain what the hidden state means physically, and none test passivity, meaning the guarantee that the model cannot create energy.

The method is borrowed from MC-LSTM. That paper restructures the LSTM cell so that mass is conserved by construction and reports better long-horizon behavior, for example a damped-pendulum MSE of 0.01 against 0.07 for a standard LSTM. We carry the same idea from mass conservation over to energy dissipation. A further constraint comes from MC-LSTM's own reported cost: its forward pass scales as $O(K^3)$ in the number of memory cells $K$, which is too slow for 1 kHz. Our cell therefore uses diagonal, per-branch recurrence, which costs $O(K)$ per step.

SurgiSim is the closest physics-side work. It fits a Maxwell viscoelastic model through differentiable material-point simulation, reconstructed from endoscopic video. It recovers bulk parameters offline and is not a real-time force model. We aim at the online force map that a haptic loop needs.

Core Claim

Let $u_t$ be the tool's indentation, measured relative to the tissue surface, at step $t$, and let $\Delta u_t = u_t - u_{t-1}$. The cell keeps $K$ branch stresses $\sigma^{(k)}_t$ and outputs force $F_t$:

$$\sigma^{(k)}_t = \alpha_k(u_t)\,\sigma^{(k)}_{t-1} + g_k(u_t)\,\Delta u_t, \qquad F_t = f_e(u_t) + \sum_{k=1}^{K}\sigma^{(k)}_t .$$

Here $\alpha_k \in (0,1)$ is a decay gate that sets branch $k$'s relaxation time. The branch stiffness is $g_k \ge 0$, and $f_e$ is a monotone elastic backbone. All three are small networks of the current deformation and optional context features, such as tissue type or visual features in the style of the RCNN. With constant $\alpha_k$ and $g_k$, this is exactly a discretized generalized Maxwell model, which dissipates energy whenever its parameters are positive. With deformation-dependent coefficients it resembles a GRU whose gates are diagonal and sign-constrained.

Passivity is not automatic when the coefficients depend on deformation. We will therefore do two things. First, we will derive a sufficient condition: the branch energy $\sigma^2/(2g)$ must be non-increasing under the update. Second, we will add a cheap correction that enforces the condition by projecting $\alpha_k$ at run time when it would be violated. Empirically, we report the net work over every closed loading cycle in the test set.

The prediction is specific. This cell should match a same-size GRU in force error on rates it was trained on. It should be clearly better at unseen loading rates and on long holds, because relaxation behavior is built into its structure rather than learned. It should produce zero passivity violations where the GRU produces some.

Research Program

We first build a viscoelastic tissue force benchmark, because the indexed literature has none with rate variation and hold phases. A dVRK or linear-stage indenter with a 6-axis force sensor will load ex-vivo porcine liver, kidney and muscle. The protocols are ramp-and-hold (relaxation), triangular cycles at five speeds that mirror the 1, 3, 5, 7 and 9 mm/s grid of Experimental Analysis of Catheter Push–Pull Forces, and random-walk trajectories. Indentations are capped near the 30 mm and 7.5 N envelope of the cGNN benchmark. A second condition mounts tissue on a pulsatile pump platform similar to the one in An MR-Safe Endovascular Robotic Platform, so the tissue is cycled continuously and never sits at rest. Surface displacement is tracked optically so that $u_t$ is measured relative to the tissue surface. Silicone phantoms are tested in parallel to measure how much of the fidelity gap shows up as missing relaxation.

We then train the passive cell, an unconstrained GRU and the other baselines on identical splits. The splits hold out loading speeds, specimens, and tissue types. The cell also goes into a simulated and then a physical haptic loop with a virtual-coupling controller, where we count instabilities. Finally, we test transfer. As a temporal head on the cGNN's per-node force output, the cell checks whether memory can be added to a spatial model without losing its 3 ms budget. On the existing fiber-optic catheter dataset, it checks whether a physically structured state handles sensor relaxation as well as a free GRU.

Baselines

The naive reference is a static map from current displacement to force: an MLP of the same width, the memoryless assumption shared by cGNN and NNWarp. It defines how much error is due to history at all. A classical generalized Maxwell model (a Prony series, meaning a sum of exponential relaxation terms) with a nonlinear elastic term is fit per tissue by least squares. It is the physicist's answer and the most dangerous baseline, because it is passive and extrapolates in rate by design. Our cell must beat it in force error on nonlinear, large-strain trials, or the learned parts add nothing.

The strongest learned competitor is an unconstrained GRU of matched parameter count, given the same inputs, training budget and hyperparameter search (equal trials over learning rate and hidden size). It is strongest because it is what won on the closest existing benchmark. On the catheter shift-of-reference data, Roughly Collected Dataset for Contact Force Sensing Catheter reports GRU MAE of 2.46 g against 3.01 g for a Transformer and 3.03 g for a fully connected network. On that dataset, beating the GRU means matching 2.46 g within noise while staying passive. On our tissue benchmark it means equal in-distribution RMSE and lower RMSE at held-out speeds. For vision-plus-kinematics force estimation, the RCNN reference reports Pearson correlation on the vertical force of 0.8957 for pushing and 0.7164 for pulling on its dataset. We rerun it as a baseline on our phantom recordings and do not compare numbers across datasets. For rate-dependent ex-vivo tissue force, no reported result was found on Lacuna.

We also run a fixed-window LSTM with window 4, following Efficiently Calibrating Cable-Driven Surgical Robots, and a Transformer over a fixed history. Two ablations isolate the key ingredient. The first keeps the diagonal branch structure but drops the sign and range constraints, which separates the effect of the architecture from the effect of the constraint. The second drops the passivity projection.

Why not just…?

Why not just fit a classical viscoelastic model? This is a real threat. If porcine tissue under these loads is close to a linear-in-time model with a fixed elastic curve, a few fitted constants will match us and should be preferred. We expect it to fail where relaxation times depend on strain and where the elastic backbone stiffens sharply, which is common for soft tissue at large indentation. The held-out-strain trials settle the question directly. If the Prony fit is within noise of our cell, we report that and recommend the classical model.

Why not just use a GRU and clip the output force? Clipping limits how large the force gets, not where the energy goes. A model can stay within force limits and still return more work than it absorbed on each cycle, which is the source of haptic instability. We measure net work per closed cycle and instability counts in the haptic loop for both models. If the GRU turns out to be empirically passive on all test data, the remaining argument is extrapolation at held-out speeds, and that is tested separately.

Why not just a longer fixed window, as in the dVRK calibration work? A window of 4 suited cable slack. Tissue relaxation lasts seconds, which is thousands of steps at 1 kHz, so a window long enough to cover it would be large and slow. The ramp-and-hold trials show whether windowed models track the relaxation tail. We include windows long enough to give them a fair chance at a matched latency budget.

Proposed Experiments and Decisive Checks

The decisive experiment is speed extrapolation: train on 1, 3 and 5 mm/s and test on 7 and 9 mm/s, using the ex-vivo tissue benchmark. The passive cell should keep its error close to the in-distribution level while the GRU's error grows. If both degrade equally, the core claim fails. A second check is long-hold relaxation error, measured as integrated error over a 30-second hold. A third is the passivity audit: net generated energy over all closed cycles, plus the instability rate in the physical haptic loop at a 1 kHz update with under 1 ms inference.

On the pulsatile platform we test whether memory of continuous cycling improves force prediction over the static map, which addresses the static-organ assumption in the RCNN study. On the catheter dataset we report MAE against the 2.46 g GRU figure. On the cGNN indentation benchmark we confirm that adding the temporal head does not hurt the static 0.37 N force error and that the per-step latency stays near 3 ms. Silicone-versus-tissue comparisons of the fitted relaxation times give a direct, quantitative measure of how much of the bio-mechanical fidelity gap lies in time-dependent behavior that current phantoms do not have.