Soft and continuum robots (snake-like arms, catheters, tentacle manipulators) have no rigid joints. Their shape is usually summarized by the piecewise constant curvature (PCC) assumption: each section of the body is treated as a perfect circular arc. The assumption is cheap, closed-form, and everywhere in the literature. Its reported accuracy, however, cannot be compared across papers. One lab reports a standard deviation in millimetres on a single planar segment, another reports tip error as a percentage of robot length on a concentric-tube robot, and a third reports locomotion speed error. None of them states how far the shapes in its test data actually were from circular arcs. A model evaluated only on gently, uniformly bent shapes will look excellent. The same model evaluated on shapes sagging under gravity or twisted by contact will look poor. Neither number says anything about the model itself unless the data regime is disclosed.
We propose to make that regime explicit and measurable. The idea has two parts. First, every dataset of continuum robot shapes should be accompanied by a small set of computable descriptors: how non-uniform the curvature is within each section, how often the body is nearly straight, and how much it twists. Together these form a "deformation regime card." Second, every shape model should be scored by its excess error above the best error any PCC-family shape could achieve on that same data, reported separately for each regime. We will test whether this reporting changes conclusions. The specific claim is that model rankings between PCC, its singularity-free rigid-link variant, and rod-mechanics estimators flip across regimes. If they do, the single aggregate numbers now reported cannot rank these models.
Core Claim
The accuracy of a low-order continuum shape model is mostly a property of the test distribution, not of the model. It can be predicted from three data descriptors. Define, for a ground-truth backbone shape $x$ split into $n$ sections, the PCC floor
$$e^*_{n}(x) = \min_{\theta \in \Theta_n} \frac{1}{L}\int_0^L \lVert r_\theta(s) - r_x(s) \rVert \, ds,$$
where $r_x(s)$ is the true backbone position at arc length $s$, $L$ is the robot length, $\Theta_n$ is the set of all $n$-arc shapes, and $r_\theta$ is the backbone of one such shape. The floor is the error a perfect PCC estimator would still make, normalized by length so it is comparable across robots. The three descriptors are the following. The curvature non-uniformity index is the within-section variance of curvature divided by its mean square. The straightness fraction is the share of sections with curvature below a threshold, which is where the standard arc parameterization divides by near-zero curvature. The torsion index is the integrated twist per unit length. We will report an estimator's excess error $e - e^*_n$ in bins of these descriptors, next to the usual aggregate. Our hypothesis is that an estimator's advantage depends on the regime:
- The algebraic rigid-link equivalent model dominates in the near-straight bin.
- Rod-mechanics estimators dominate in the high non-uniformity and torsion bins.
- In the low non-uniformity, moderate-curvature bin, the models are indistinguishable.
If the hypothesis holds, most of the published gaps between models reflect which regime each test set happened to sample.
Research Program
The work builds on three kinds of prior results.
Two papers motivate it by showing regime-dependent behaviour without naming it. Approximate Piecewise Constant Curvature Equivalent Model and Their Application to Continuum Robot Configuration Estimation replaces each arc with two rigid links and five revolute joints (the "2L-5R" model) to remove the zero-curvature singularity. It reports its gain on a single planar segment that includes the straight configuration. That is exactly the regime where one would expect the largest gain, and the paper explicitly leaves 3D bending, torsion, and multi-segment accumulation open. Continuum Robot State Estimation with Actuation Uncertainty instead estimates a discretized Cosserat rod through a probabilistic factor graph; a Cosserat rod is a beam model that tracks bending, twist, shear, and stretch along the backbone. It notes that tip error in millimetres hides poorly conditioned axial and force errors, and it never compares against PCC. Our proposal puts both methods on identical data and asks where each one wins.
Two further works supply methods we reuse. Elastica: A compliant mechanics environment for soft robotic control provides open-source Cosserat rod simulation, which yields dense ground-truth backbones under controlled gravity, tip loads, distributed contact, and torsion. A Neural Network-based Framework for Fast and Smooth Posture Reconstruction of a Soft Continuum Arm reconstructs Cosserat posture from sparse markers through a low-dimensional strain basis, and it serves as the learned competitor.
Two sources constrain the design. Reproducibility of deep learning in digital pathology whole slide image analysis showed that undisclosed data-preparation choices caused replications to collapse; for example, one model's slide accuracy fell from 92% to 54.7%. We take its lesson that disclosure must be concrete and checkable. We extend it in one respect: its checklist covers how data were processed, not which part of the problem space the test data cover. ManiSoft: Towards Vision-Language Manipulation for Soft Continuum Robotics is the closest shared benchmark. It is public and Cosserat-based, but it evaluates only end-to-end neural policies, and it attributes their failures to a "proprioception bottleneck": the policies cannot infer the arm's shape from pixels.
We will build four things.
The first is an open generator of Elastica datasets that sweeps load conditions so the three descriptors cover their ranges deliberately rather than by accident. It will record the seeds, material parameters, contact parameters, and actuation profiles for every trajectory.
The second is a solver for the floor $e^*_n$. It minimizes over arc parameters with multi-start optimization, and in near-straight sections it uses the 2L-5R parameterization so the floor itself does not suffer from the singularity.
The third is regime-stratified splits. Test bins are held out entirely from calibration data, so that no estimator is tuned on the regime it is scored on. This targets the in-distribution leakage problem flagged for kinematic benchmarks.
The fourth is a small physical dataset from a tendon-driven rod tracked by an electromagnetic sensor. It will be released with its regime card, raw sensor streams, calibration, and actuator latency logs, so that the simulated conclusions are checked on hardware.
We will also compute regime cards for the ManiSoft trajectories. Because ManiSoft is Cosserat-based, the full backbone of every trajectory is available.
Baselines
The naive reference is a length-matched straight rod pointing along the base axis, plus a mean-shape predictor fit on the calibration split. Any estimator that fails to beat these within a bin is uninformative there.
The first real baseline is standard three-parameter PCC fit to the same sparse observations, which is what most applied papers use. It will be run through the same least-squares fitting code as the other geometric models, from the same marker positions.
The strongest geometric competitor is the 2L-5R model of Cheng et al.. On its own planar hardware that paper reports an error standard deviation of 0.558 mm, against 3.468 mm for standard PCC. We will reimplement it from the paper, and the reimplementation itself is an instance of the transparency question.
The strongest mechanics competitor is the midpoint-discretized Cosserat factor graph of Ferguson et al.. It reports 1.95 mm mean tip error and 0.49 N force error on the Virtuoso robot, and 1.3% of robot length tip error in simulation, with 18 to 25 ms solve times. We will run it with the same ten-node discretization the paper found sufficient.
The learned competitor is the strain-basis network of Wang et al., which reports below 0.2% simulated error. It is trained only on the calibration regimes.
Fairness rules apply to all methods. Every method sees the same marker subsets and the same noise. Each gets the same tuning budget, measured in hyperparameter trials on the calibration split. Latency is timed on one CPU.
"Beating" a method has a specific meaning here. It does not mean our method wins; the proposal contributes an evaluation, not an estimator. It means the regime-stratified report produces at least one statistically significant ranking reversal, under bootstrap confidence intervals over trajectories, that the aggregate length-normalized error does not reveal.
Two ablations isolate the key ingredient. The first replaces our three descriptors with the obvious alternatives, raw load magnitude and random bins, to test whether the descriptors specifically carry the predictive power. The second compares stratified splits against IID (random, same-distribution) splits to measure how much leakage inflates reported accuracy.
Why not just…?
Why not just report error normalized by robot length on a shared dataset? This is the most serious alternative. Ferguson et al. already use percentage-of-length error, and it is simpler than regime cards. It fixes cross-robot scaling but not cross-distribution mixing. A shared dataset dominated by moderate uniform bends would still rank models by their moderate-bend behaviour. The decisive test is direct. If aggregate length-normalized error gives the same model ranking in every descriptor bin, then the cards add nothing and we will say so. We will then recommend the simpler metric plus the released generator.
Why not just use Cosserat rods everywhere and drop PCC? Cosserat estimators cost 18 to 25 ms per solve in Ferguson et al., against milliseconds for the algebraic solver in Cheng et al.. That gap matters for planners such as Efficient RRT*-based Safety-Constrained Motion Planning for Continuum Robots in Dynamic Environments, which call the kinematics thousands of times. The practical question is therefore when the cheap model is good enough, and the floor answers it: wherever $e^*_n$ is below the task tolerance, PCC loses nothing.
Why not just release the raw data with a Fell-style checklist? The checklist of Fell et al. would fix the fact that no dataset in Section 2 of the literature is openly released, and we adopt it. But a fully documented dataset that samples only easy shapes still overstates model quality. Coverage of the problem space is a separate disclosure. The leakage ablation measures how large that separate effect is.
Why not just learn shape end to end and skip kinematic models? On ManiSoft, Diffusion Policy fell from 31.6% success in clean scenes to 2.2% in randomized ones (ManiSoft), and the authors blamed shape inference. Our ManiSoft regime cards test whether an explicit PCC state is even a plausible remedy. If most ManiSoft arm shapes have a low floor, then an arc-based observer could plausibly supply the missing shape information. If the floor is high, it cannot.
Proposed Experiments
Experiment 1: regime sweep. On the Elastica sweep, we compute the floor and the excess error of all five estimators across a grid of the three descriptors. We use 2, 3, and 5 sections and marker counts from 3 to 10.
Experiment 2: sufficiency of the descriptors. We fit a simple regression from the three descriptors to each estimator's error. Our claim is supported only if this regression explains most of the variance and clearly outperforms the load-magnitude ablation. We report it on held-out bins.
Experiment 3: reversals. We test for ranking reversals between the 2L-5R model and the Cosserat factor graph.
- Prediction 1: the 2L-5R model wins in the straightness bin.
- Prediction 2: the Cosserat factor graph wins in the high-torsion bin.
- Falsification: no significant reversal appears in any bin.
Experiment 4: leakage. We compare errors under IID splits against regime-held-out splits for the learned estimator, which is the method most exposed to leakage.
Experiment 5: hardware check. We repeat Experiment 3 on the physical tendon rod, to check that the simulated ranking pattern holds on hardware.
Experiment 6: ManiSoft cards. We publish regime cards for ManiSoft. As one bounded downstream check, we feed a pixel-fitted PCC state to ManiSoft's Diffusion Policy baseline in randomized scenes and compare against the 2.2% reported there.
Link between Experiments 1 and 6. The PCC floor on those scenes should predict whether this helps. Specifically, a gain should appear only in trajectories whose floor is below the object-placement tolerance.
References
- Approximate Piecewise Constant Curvature Equivalent Model and Their Application to Continuum Robot Configuration Estimation
- Continuum Robot State Estimation with Actuation Uncertainty
- Elastica: A Compliant Mechanics Environment for Soft Robotic Control
- A Neural Network-Based Framework for Fast and Smooth Posture Reconstruction of a Soft Continuum Arm
- Reproducibility of deep learning in digital pathology whole slide image analysis
- ManiSoft: Towards Vision-Language Manipulation for Soft Continuum Robotics
- Efficient RRT*-based Safety-Constrained Motion Planning for Continuum Robots in Dynamic Environments