{"id":"6a15cc97-b892-42af-bde6-6878eaeae123","arxiv_id":"2601.04493","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A sparse factor-graph formulation of the discrete Cosserat rod with midpoint strain integration gives real-time joint estimation of shape, external loads, and actuation for tendon-driven, parallel, and concentric-tube continuum robots, with a marginal-covariance shortcut for manipulator Jacobians.","lead":"This paper builds one computer model that estimates a flexible robot arm's shape, the forces it feels, and how uncertain its motor inputs are — all together and in real time. It is demonstrated in simulation on two robot types and on a real surgical arm, and the same math conveniently hands over the steering matrix (Jacobian) used for control.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Jacobian extraction in Eq. 23 is a posterior regression coefficient, not the mechanical manipulator Jacobian; measurement factors shrink it, so the unqualified Jacobian/control claim needs correction.","rationale":"The reader's verdict is CONDITIONAL and identifies the physical fidelity of the discrete rod model as the weakest assumption. I agree that calibration and discretization deserve scrutiny, but the sharper, more testable weakness is the Jacobian extraction. The paper's central contribution is not only estimation accuracy but also 'direct access to manipulator Jacobians through the linearized factor graph.' The formula in Eq. (23) is a linear regression coefficient in the posterior distribution, and the posterior includes measurement factors. As shown in a simple linear-Gaussian calculation, such a coefficient is attenuated by measurement noise and therefore is not the standard quasi-static manipulator Jacobian. The paper gives no derivation of why this posterior slope should equal the mechanical Jacobian, and it does not disclose the graph configuration used when extracting the Jacobian for control. The trajectory-tracking results may still be valid if the method is interpreted as a closed-loop estimation Jacobian, but the claim as stated is misleading. This strengthens the need for the conditional verdict: the framework is promising, but the Jacobian claim requires a concrete correction and a specification of which factors are active. I therefore keep the verdict UNCHANGED while elevating this specific issue for the authors to address.","tokens_in":11987,"tokens_out":10364,"duration_ms":121318,"concrete_test":"In the Sec. IV-C tracking setup, compute J from Eq. (23) using the full posterior Hessian. Then recompute J from the same graph with the tip-position factor's information matrix set to zero (or equivalently its covariance set to infinity), and also compute a finite-difference Jacobian of tip pose with respect to commanded tensions in the mechanics/actuation-only model. If the full-posterior J differs from the model-only J by more than 10%, Eq. (23) does not return the manipulator Jacobian; the paper should either extract Jacobians from the model-only graph or explicitly define them as posterior estimation Jacobians.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is the claimed Jacobian extraction in Sec. IV-A. Eq. (23) defines J = Σ_{T_K Q} Σ_{QQ}^{-1}, the slope of the posterior conditional expectation E[ξ_K | Q] computed from the full linearized factor graph. But the tracking experiments include the tip-position measurement factor (22) and the actuation prior (18). Because the pose block of the posterior Hessian contains measurement information, the Schur-complement slope is -H_{T_K T_K}^{-1} H_{T_K Q}, which is attenuated by measurement noise. In a linear scalar proxy T = aQ + noise with tip measurement noise σ_z, this expression gives a/(1 + σ_T^2 / σ_z^2), not a. Thus J is an estimation/feedback gain, not the mechanical manipulator Jacobian mapping commanded actuation to tip motion under fixed external loads. The paper uses this Jacobian for trajectory tracking and claims 'direct access to manipulator Jacobians,' but it never specifies which factors are active when extracting J or compares with finite-difference Jacobians of the open-loop model. In the open-loop benchmark (Sec. IV-B) no tip measurements are used, while tracking (Sec. IV-C) adds them, so the Jacobian used for control is likely measurement-contaminated. This does not falsify the estimation framework, but it does invalidate the unqualified 'manipulator Jacobian' claim and should be corrected or re-labeled.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a discrete Cosserat rod formulation for continuum robot state estimation, cast as a sparse factor graph on SE(3). The kinematics use a midpoint strain integration rule (Eqs. 7-8), and the mechanics are discretized by concentrating external wrenches at arclength nodes (Eqs. 9-12). Actuation is introduced as random variables with priors (Eq. 18) and tendon actuation factors (Eqs. 19-21); parallel continuum robots are modeled by connecting multiple rod graphs with platform constraints (Eqs. 24-25). The authors claim high numerical accuracy with few nodes (1% error at K=10, 0.1% with K<30), real-time solve times (2-24 ms), direct extraction of manipulator Jacobians from the linearized graph (Eq. 23), and experimental validation on a concentric tube robot with 1.95 mm tip position accuracy and 0.49 N tip force accuracy. They also disclose that axial z-force estimation is ill-conditioned (Sec. VI).","tokens_in":12274,"tokens_out":6884,"duration_ms":66943,"significance":"If the claims hold, the framework is a useful contribution: it extends prior SE(3) / GP-Cosserat estimators by explicitly modeling actuation uncertainty, covers tendon-driven, parallel, and concentric-tube architectures in one graph, and demonstrates real-time sparse inference. Strengths include the internal consistency of the core error models with the stated SE(3) conventions, the BVP comparisons that quantify discretization error, the real physical experiment on a concentric tube robot, and the honest disclosure of the z-force degeneracy. The main weaknesses are that the 'manipulator Jacobian' extraction in Eq. (23) is a posterior regression coefficient rather than the mechanical Jacobian when measurement factors are active, and that the experimental calibration lacks identifiability analysis. These issues are local and correctable, but they affect load-bearing claims in the abstract and in the control experiments.","major_comments":[{"comment":"Eq. (23) defines J_TKQ = Sigma_TKQ Sigma_QQ^{-1}, the best linear predictor of tip pose from actuation under the posterior distribution that includes the tip-position measurement factor (22) and the actuation prior (18). This is an estimation/feedback gain, not the mechanical manipulator Jacobian mapping commanded actuation to tip motion under fixed external loads. Measurement information attenuates the slope; in a scalar proxy T = aQ + w with measurement z = T + v, the posterior slope is a sigma_v^2 / (sigma_w^2 + sigma_v^2), not a. Since Sec. IV-B uses no tip measurements while Sec. IV-C tracks with them, the Jacobian used for control is measurement-contaminated. Please re-label Eq. (23) as a posterior feedback gain or specify that measurement factors must be removed when extracting the mechanical Jacobian, and compare against finite-difference Jacobians of the open-loop model.","section":"Sec. IV-A, Eq. (23)"},{"comment":"The experimental calibration performs a joint maximum-likelihood optimization over tracker pose, tube curvatures, tube bending/torsion stiffness, and all actuation noise parameters from a single calibration dataset. No identifiability analysis, parameter count, or cross-validation is reported. If these parameters are non-identifiable, the calibrated values can absorb modeling error, and the reported 1.95 mm / 0.49 N accuracy and the uncertainty envelopes may not generalize. Please report parameter counts, constraints, cross-validation, and the sensitivity of downstream estimates to calibration choices.","section":"Sec. VI"},{"comment":"The simulation benchmarks compare the MAP solution against deterministic BVP solvers of the same Cosserat equations, so they validate the discretization and optimization, not the physical model fidelity. The discrete-load approximation in Eq. (9) is exact only for point loads such as tendon discs; for distributed loads, accuracy depends on node count K, and no convergence study or experiment with distributed contact loading is provided. The claims of validation 'across multiple robot architectures' should be tempered, or supplemented with a test against an independent model or an experiment with distributed loading.","section":"Sec. III-B / IV-B / V-A"}],"minor_comments":[{"comment":"Typography: 'realtive' should be 'relative', and the trailing phrase 'mean force error).' is incomplete/duplicated.","section":"Fig. 10 caption"},{"comment":"The text refers to 'the midpoint noise model (7)', but Eq. (7) is a deterministic strain rule; the noise enters later through n_epsilon_k. Please correct the terminology.","section":"Sec. III-E"},{"comment":"The linearization of the actuation error e_Dd in Eq. (21) is deferred to 'source code', which is not available at review. Since the source is only promised 'upon acceptance', please include the full derivatives in the appendix or in a supplementary document.","section":"Appendix I"},{"comment":"The z-force degeneracy is an important limitation and is mentioned only in passing at the end of Sec. VI. It should be stated in the abstract or conclusions with a quantitative comparison of prior vs. posterior variance, since it tempers the force-sensing claims.","section":"Sec. VI"}],"recommendation":"major_revision","confidential_remarks":"The central factor-graph formulation appears sound and the experimental work is valuable. The main issue is the unqualified 'manipulator Jacobian' labeling in Eq. (23), which is a posterior feedback gain when measurements are active; this is fixable by re-labeling or by specifying the factor set used for extraction. I would also like to see the calibration identifiability analysis before acceptance, as it underpins the experimental claims. The paper is within the journal's scope and should be publishable after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does something genuinely new: it discretizes the Cosserat rod equations with a midpoint strain rule into a sparse factor graph, puts actuation variables in the state, and shows the same graph handles tendon-driven, parallel, and concentric-tube robots. The accuracy comparisons against BVP solvers are honest and convincing — 1.3% and 0.06% tip error over trajectories is real evidence for the discretization. The experimental setup on the concentric tube robot is a genuine hardware validation, and the authors disclose the z-force identifiability problem rather than burying it.\n\nThe biggest issue is the Jacobian extraction in Eq. 23. Calling it the manipulator Jacobian is misleading. The expression Σ_{T_K Q} Σ_{QQ}^{-1} is the slope of the posterior conditional expectation E[ξ_K | Q]. When tip measurement factors are active, that slope is attenuated by measurement noise — it is a feedback gain, not the mechanical map from actuation to tip motion under fixed loads. The paper uses this Jacobian for trajectory tracking in Sec. IV-C, where the measurement factors are active, so the control Jacobian is measurement-contaminated. This doesn't break the estimation framework, but the unqualified 'direct access to manipulator Jacobians' claim needs to be corrected or relabeled.\n\nSecond, the experimental section reports only means — no variances, trial counts, or hardware timing details. Third, code and data are promised 'upon acceptance'; for reproducibility that should be released in the open, and reviewers should ask for it. Fourth, the calibration fits a bundle of parameters (tracker pose, curvatures, stiffnesses, noise covariances) with no identifiability analysis; that's worth a closer look, though the held-out evaluation mitigates some concern.\n\nThe core formulation, however, is sound. The error models are self-consistent, the sparse optimization is real, and the extension to parallel robots via platform constraints is natural and demonstrated. I'd send this to peer review — likely with requests for the code release, more detailed experimental data, and a reworking of the Jacobian claim.","headline":"Solid factor-graph framework for continuum robot state estimation; the Jacobian-extraction claim is oversold and the experimental reporting needs detail, but the core formulation is new and worth refereeing.","tokens_in":12883,"tokens_out":2428,"would_cite":true,"duration_ms":24602,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A discrete Cosserat rod factor graph with midpoint strain integration can jointly estimate shape, external loads, and actuation inputs of continuum robots in real time, while yielding manipulator Jacobians from the same linearized graph.","keywords":["continuum robots","state estimation","Cosserat rod","factor graphs","actuation uncertainty","tendon-driven robots","concentric tube robots","force sensing"],"falsifier":"A comparison experiment on a robot with a known distributed load (continuous contact along the shaft rather than point contact), with dense force sensors along the backbone: if estimated shape and force bias systematically with node count even at fine discretization, the midpoint-strain/discrete-load model is the culprit. Also, a calibration identifiability study: perturb stiffness and noise parameters and check whether the joint maximum-likelihood estimate recovers them uniquely; if not, the uncertainty envelopes are not trustworthy.","tokens_in":11731,"feed_emoji":"🤖","tokens_out":4447,"duration_ms":39334,"temperature":0.7,"pith_summary":"This paper argues that continuum robots — thin flexible arms used in surgery — can be tracked in real time by reformulating the Cosserat rod equations as a discrete factor graph, with actuation treated as an uncertain variable rather than a fixed input. The key move is a midpoint strain integration rule that gives high numerical accuracy with only about 10–30 nodes along the backbone, keeping the optimization sparse. From the same linearized graph, the paper extracts manipulator Jacobians, so the estimated state can drive trajectory tracking. The authors show the approach on tendon-driven and parallel robots in simulation and on a surgical concentric tube robot, reporting mean tip position error of 1.95 mm and mean tip force error of 0.49 N at millisecond solve times.","feed_headline":"One factor graph estimates shape, load, and actuation in real time","feed_subtitle":"Midpoint strain integration reaches 0.1% error with under 30 nodes, and the same graph yields control Jacobians.","key_machinery":"The central object is a discrete Cosserat rod factor graph: the backbone is divided into arclength nodes, each holding pose, stress, and wrench variables, connected by kinematics factors (midpoint strain rule) and mechanics factors (stress propagation with point loads). The midpoint strain rule, averaging endpoint stresses over each interval, is what buys accuracy with few nodes. A second mechanism is the actuation factor linking tendon tensions to backbone wrenches, plus platform constraint factors connecting multiple rods for parallel robots. The linearized graph's covariance blocks give the Jacobian.","core_discovery":"The paper's central discovery is that the discrete Cosserat rod, when integrated with a midpoint strain rule, induces a sparse factor graph over SE(3) poses, internal stresses, and external wrenches; adding actuation variables as random nodes lets the graph jointly infer shape, external loads, and actuation inputs under uncertainty. A corollary is that the posterior covariance at the optimum directly yields the manipulator Jacobian relating actuation to tip motion, avoiding finite differences or separate sensitivity integration. The experiments show that this formulation matches a boundary-value-problem solver to better than 1.3% of robot length in open-loop simulation, and in hardware achie","pith_inferences":["A natural extension, not drawn in the paper: the factor-graph structure should also support nonlinear constitutive laws, since the midpoint strain factor does not depend on linear stiffness; repeating the accuracy benchmark for nonlinear materials would test this.","The observed axial-force ill-conditioning suggests a testable design rule: a second sensing modality along the shaft, not just tip pose, may be needed to make 3D force sensing fully observable; the paper itself flags the z-direction weakness.","Because the Jacobian comes from posterior covariances, the same linearization could be used for information-driven palpation, selecting future contacts to minimize force-estimate uncertainty—an application the paper does not mention.","A follow-up validation not in the paper: leave-one-out calibration checks would reveal whether the jointly calibrated stiffness and noise parameters are identifiable or merely absorbing model error; if the latter, the reported uncertainty envelopes could be overconfident."],"forward_implications":["Real-time joint shape/load/actuation estimation becomes practical for tendon-driven, parallel, and concentric-tube continuum robots, with reported solve times around 2–24 ms.","Manipulator Jacobians come for free from the linearized factor graph, enabling closed-loop trajectory tracking without finite differences.","The same graph can run forward (known loads to shape) or inverse (observed shape to loads) by changing which priors and measurements are attached.","Actuation uncertainty, including backlash and torsion in endoscopic drive channels, is incorporated as noise on base poses rather than ignored.","Point-load discretization makes the model exact for tendon-disc loads and approximate for distributed loads, with accuracy controlled by node count."],"fun_headline_variants":["Sparse factor graph jointly estimates shape, load, and actuation","One graph, real-time estimation of shape, load, and actuation under uncertainty","Actuation uncertainty resolved: joint shape, load, and input estimation in real time","Factor graph over SE(3) yields shape, loads, actuation, and Jacobians live","Midpoint Cosserat factor graph: real-time concurrent shape, load, and actuation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the discretized rod model—constant strain per interval computed from averaged endpoint stresses, loads concentrated at nodes, and a linear stiffness law—faithfully represents the physical robot, so any bias from distributed loads or calibration absorbing model error cannot be corrected by the factor graph.","fun_headline_variants_meta":{"raw":{"variants":["Sparse factor graph jointly estimates shape, load, and actuation","One graph, real-time estimation of shape, load, and actuation under uncertainty","Actuation uncertainty resolved: joint shape, load, and input estimation in real time","Factor graph over SE(3) yields shape, loads, actuation, and Jacobians live","Midpoint Cosserat factor graph: real-time concurrent shape, load, and actuation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1141,"prompt_tokens":664,"completion_tokens":477,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":408,"completion_tokens_details":{"reasoning_tokens":383}},"tokens_in":408,"tokens_out":477,"duration_ms":4438,"temperature":1.0,"reasoning_tokens":383,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T12:02:49.463340+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A comparison experiment on a robot with a known distributed load (continuous contact along the shaft rather than point contact), with dense force sensors along the backbone: if estimated shape and force bias systematically with node count even at fine discretization, the midpoint-strain/discrete-load model is the culprit. Also, a calibration identifiability study: perturb stiffness and noise parameters and check whether the joint maximum-likelihood estimate recovers them uniquely; if not, the uncertainty envelopes are not trustworthy.","supporting_citations":[],"review_version":1}