{"id":"da1b2e33-27d5-4b42-95ad-305e2e12824f","arxiv_id":"2603.29383","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An IMM filter whose modes differ only in foot-velocity noise, combined with a rolling-contact measurement model, cuts drift in IMU/encoder-only quadruped odometry by 30-70% in simulation and ~55% on a real robot.","lead":"A state-estimation method for legged robots tracks the robot's motion from internal sensors alone by treating foot-ground contact as a rolling touch, and by probabilistically switching between two contact models when the foot starts to slip. On simulations and a Unitree AlienGo quadruped it reports roughly one-third to one-half the position drift of a strong invariant-Kalman baseline, at about 1 ms per update.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Slip mode's enlarged Q_vf strengthens, not weakens, the rolling pseudo-measurement in Eq. (16), inverting the stated mechanism for IMM-PO's slippery-terrain gains.","rationale":"The reader correctly identified the rolling-contact model and the untested slip mechanism as the weakest point, but I found a more specific and internally checkable problem: the only stated difference between modes (Q_vf) has the opposite effect on the rolling pseudo-measurement to what the paper claims. This does not by itself overturn the empirical tables, which could still be reproducible, but it removes the paper's causal account of why IMM-PO works under slip. The issue is concrete, local to Eqs. (16) and Remark 2, and directly testable. Because the empirical claims remain unverified (no code, no parameters, single runs) and the mechanism is now suspect, the existing CONDITIONAL verdict is appropriate, but the condition should include a check of this Q_vf/R_k inversion rather than only reproducibility concerns.","tokens_in":15304,"tokens_out":9503,"duration_ms":115903,"concrete_test":"Run the two-mode IMM-PO exactly as specified by Eqs. (9)-(23) in a simulated slip episode where the true foot velocity ramps away from ω_f × r, with α_2 = 100 and all other parameters fixed. Compare the posterior \\v_f and the slip-mode likelihood against the α_2 = 1 single-mode filter. The paper's mechanism predicts the slip mode should track the deviating foot velocity; the standard Kalman-gain argument predicts it will instead pull v_f closer to ω_f × r. If the latter occurs, the stated mechanism is inverted and the reported slippery-terrain gains require an alternative explanation or an unreported modification to R_k.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim attributes the slippery-terrain improvement to a slip mode that 'weakens constraint confidence' via enlarged foot-velocity process noise Q_vf (Remark 2; Sec. III-C3). But the rolling constraint is implemented as a direct pseudo-measurement in Eq. (16): the third row is v_f − (ω_f × r), with some measurement noise R_k. For a direct measurement of v_f, the Kalman gain for that row is K = P_vf H^T (H P_vf H^T + R_k)^{-1}. Increasing Q_vf inflates the predicted covariance P_vf, driving K toward the full-information value and pulling the posterior v_f more tightly onto the rolling value ω_f × r. The slip mode therefore hardens the rolling constraint, not weakens it. This contradicts the paper's explanation of Fig. 8 and the ablation in Table III. If the empirical gains are real, they cannot be caused by the stated Q_vf-only mechanism; they would have to come from IMM covariance mixing, the added foot-velocity state acting as a low-pass filter, or an unreported mode-dependent change in R_k. That missing distinction is load-bearing because the paper's claimed contribution is precisely the rolling-aware, mode-adaptive design.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an IMM-based proprioceptive odometry (IMM-PO) for legged robots. The state is augmented with foot velocities, and a rolling-contact constraint v_f = ω_f × r is used as a pseudo-measurement. Two contact modes (nominal rolling and slippery) are fused via the IMM, with modes differing only in the foot-velocity process noise Q_vf. Experiments in Gazebo and on an AlienGo robot compare IMM-PO against KF, ESKF, IEKF, ESKF-VB, KalmanNet, and a single-model rolling ESKF-R, reporting consistent improvements in ATE and RPE on flat, slippery, uneven, and slope scenes. The paper also includes an ablation study and a detailed KalmanNet failure analysis.","tokens_in":15552,"tokens_out":4974,"duration_ms":53156,"significance":"If the reported results are robust, the paper makes a useful contribution to proprioceptive odometry: it is the first to combine an explicit rolling-contact model with IMM-based mode adaptation for legged robots, and the reported gains over well-established baselines are substantial (e.g., on real complex terrain ATEpos 0.102 m vs 0.227 m for IEKF). The ESKF/IMM mathematics are standard and internally consistent, and the headline percentages check out against Table I. However, the significance is conditional on resolving the mechanism issue and on providing statistical support for the experimental claims; as written, the central explanatory claim appears to be contradicted by the filter equations.","major_comments":[{"comment":"The stated mechanism for the slip mode is inverted. Remark 2 claims that enlarging Q_vf 'weakens constraint confidence,' but in Eq. (16) the rolling constraint is enforced as a direct pseudo-measurement on v_f. For a direct measurement, the Kalman gain K = P_vf H^T (H P_vf H^T + R_k)^{-1} increases as Q_vf (and hence predicted P_vf) increases, pulling the posterior v_f more tightly onto ω_f × r. Thus the slip mode hardens, not weakens, the rolling constraint. This contradicts the explanation of Fig. 8 and the interpretation of the ablation in Table III. The authors should either change the design (e.g., use mode-dependent R_k for the rolling pseudomeasurement) or provide an alternative analysis showing how a larger Q_vf can produce the claimed effect.","section":"§III-C3 (Eq. 16) and Remark 2"},{"comment":"All quantitative results are from single runs with no error bars, no standard deviations, and no repeated trials. The abstract and conclusions claim 'superior pose estimation accuracy' over state-of-the-art methods, which requires evidence that the improvements are not due to chance or a favorable run. In simulation, multiple seeds are inexpensive; for real-world data, at least several trials should be reported. Please provide mean ± std and, where appropriate, statistical significance tests.","section":"§IV-A, Tables I–III"},{"comment":"The method's performance depends on several reported-but-not-quantified parameters: the slip-mode noise scale α_2, the Markov transition matrix Π, the foot radius vector r, the process noise Q_vf, and measurement noise R_k. None of these are given numerically, and no sensitivity analysis is presented. Since the central advantage of IMM-PO is attributed to this mode adaptation, the absence of parameter values and robustness analysis makes the empirical contribution difficult to reproduce and assess.","section":"§IV-C and §III-C3"},{"comment":"On uneven terrain, the single-model rolling filter ESKF-R is worse than the point-contact IEKF (ATEpos 0.7595 vs 0.6513), yet IMM-PO is dramatically better (0.2720). This is surprising given Remark 1's own caveat that the rolling model is less accurate on uneven terrain. The paper does not explain why adding an IMM to a worse underlying model produces such a large improvement. Without a plausible mechanism (or a sensitivity analysis over the IMM parameters), this result is not convincing and suggests the improvement may stem from an unintended effect such as the extra foot-velocity state acting as a low-pass filter.","section":"§IV-A, Table I (Uneven Terrain)"}],"minor_comments":[{"comment":"The text says 'The training details and error analysis of KalmanNet are provided in []' — the citation is empty. Since Sec. IV-D contains this material, the reference should be to that section or to a numbered citation.","section":"§IV (beginning of section)"},{"comment":"The sentence 'The robot navigated in the indoor flat surface and complex terrain, as indicated in Fig. 3' appears to reference the wrong figure; the real experiments are shown in Fig. 6, not Fig. 3.","section":"§IV-B"},{"comment":"The KalmanNet analysis is extensive and somewhat tangential to the paper's main contribution. Consider moving it to an appendix or condensing it, as it interrupts the flow of the experimental section.","section":"§IV-D"},{"comment":"Table IV shows IMM-PO has a runtime of 1.00 ms versus 0.49 ms for ESKF-R — a two-fold increase. Calling this 'comparable computational efficiency' is optimistic, even though both are within real-time limits.","section":"Table IV and §IV-C"},{"comment":"The state dimension is stated as 39, but this is not explicitly derived. Adding a short dimension count (p, v, G, 4 foot positions, 4 foot velocities, ba, bω) would help the reader verify the formulation.","section":"§III-B"}],"recommendation":"major_revision","confidential_remarks":"The mechanism issue is the most serious concern: the slip-mode design appears to enforce the rolling constraint more strongly, contrary to the paper's stated rationale. Even if the empirical numbers are reproducible, the paper's central claim about how the method works would need substantial rewriting, and the experimental evaluation needs statistical rigor. The empty reference and somewhat off-topic KalmanNet section suggest the manuscript is not yet fully polished. I would not recommend rejection if the authors can provide a corrected explanation or new experiments distinguishing the true mechanism, but the current version is not publishable in a serious journal without those changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the quick read. The paper is the first to put an IMM over contact modes for proprioceptive legged odometry, with a rolling-aware foot-velocity state. That combination is genuinely new, and the results are consistently in its favor: large ATE reductions over IEKF in simulation and a roughly 55% drop on real complex terrain. The ESKF/IMM mathematics is standard and the headline percentages in Table I check out. I'd send this to a serious referee.\n\nBut there is a real problem with the central mechanism. The paper says the slip mode 'weakens constraint confidence' by inflating the foot-velocity process noise Q_vf (Remark 2). The update, however, uses Eq. (16) as a direct pseudo-measurement of v_f against ω_f × r. In a Kalman filter, raising Q_vf inflates the predicted covariance P_vf and drives the gain for that measurement toward full information, pulling the posterior v_f harder onto the rolling value. So the slip mode hardens the rolling constraint, not weakens it. The authors have the effect backwards. This is not cosmetic: it means the reported gains, if real, are not caused by the stated adaptive-confidence mechanism. The gains might come from the extra foot-velocity state acting as a low-pass filter, from IMM covariance mixing, or from some unreported mode-dependent R_k. The paper does not test any alternative.\n\nOther soft spots are more conventional. The decisive constants (alpha_2, the Markov matrix Pi, foot radius r, Q_vf, R_k) are not reported, so the improvements could be a tuning artifact. Every number in Tables I–III is a single run with no error bars. No code or data is released, and the manuscript has an empty citation and broken figure references, so it is unfinished.\n\nWhere the paper earns credit: the survey is honest about the sparse multi-model literature, the equations are laid out cleanly, and the experiments span four simulation scenarios plus two real scenes, which is more than most proprioceptive odometry papers. The ablation includes an interesting 'no interaction' variant and a three-mode IMM, which helps bracket the design choices.\n\nBottom line: the idea is worth testing and should go to peer review, but the authors need to either fix the mechanism (e.g., inflate the measurement noise R_k in the slip mode instead of Q_vf) or provide a different explanation for the gains, report the parameters, and run repeated trials. As it stands, I wouldn't cite it yet.","headline":"A genuinely new IMM-plus-rolling combination with promising numbers, but the slip-mode mechanism is backwards as written and the tuning/experiments are underreported.","tokens_in":16193,"tokens_out":4122,"would_cite":false,"duration_ms":43493,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T40","93E11"],"pacs":[],"model":"deepseek-v4-flash","headline":"Treating a legged robot's stance foot as rolling (v_f = ω_f × r) instead of fixed, and fusing rolling and slip modes with an interacting multiple-model filter, cuts pose error by half to two-thirds on slippery and uneven ground at roughly 1","keywords":["legged robots","proprioceptive odometry","interacting multiple model","state estimation","rolling contact","slip detection","error-state Kalman filter","inertial measurement unit"],"falsifier":"Compare IMM-PO against a single-model point-contact IEKF on a soft or highly uneven terrain where rolling is not a good model; if the point-contact filter matches or beats IMM-PO, the rolling premise is not the source of the reported gains. Alternatively, remove the third measurement block in Eq. 16 (the rolling constraint on foot velocity) and keep the foot-velocity state; if accuracy is unchanged, the gains come from the state augmentation, not the rolling model.","tokens_in":15089,"feed_emoji":"🦿","tokens_out":7052,"duration_ms":68525,"temperature":0.7,"pith_summary":"This paper tackles the drift that plagues legged-robot state estimation when only proprioceptive sensors (IMU and joint encoders) are available. It argues that the common 'point foot is stationary' assumption is a major error source because real feet roll on the ground and slip on low-friction surfaces. The proposed estimator augments the state with foot velocities and constrains each stance foot to the kinematic rolling velocity ω_f × r; it then wraps this rolling-aware filter in an interacting multiple-model (IMM) framework with two modes—rolling and slip—that differ only in foot-velocity process noise. On simulation and a real quadruped, the resulting IMM-PO filter reduces absolute trajectory error by roughly half to two-thirds relative to point-contact baselines (e.g., 0.102 m vs 0.227 m on real complex terrain) while keeping runtime around 1 ms per step. The central claim is that explicitly modeling rolling motion and adapting contact confidence through IMM yields this large, consistent improvement.","feed_headline":"Two-mode rolling/slip filter halves legged-robot pose drift","feed_subtitle":"It needs only IMU and joint encoders, runs in ~1 ms, and cuts pose error to 0.1 m on real rough terrain.","key_machinery":"The central object is the rolling-aware measurement model (Eq. 16), which constrains three quantities: the relative body–foot position (via forward kinematics), the body velocity minus the foot velocity, and the foot velocity itself to the kinematic rolling velocity ω_f × r, where r is the foot radius vector. This model is embedded in two parallel error-state Kalman filters that form an interacting multiple model; the only difference between the two modes is the foot-velocity process noise covariance Q_vf^(i) = α_i Q_vf, with α=1 for rolling and α>1 for slip. The IMM mixing and fusion equations (17–23) derive mode probabilities from innovation likelihoods, enabling online switching between t","core_discovery":"The paper establishes that proprioceptive odometry for legged robots can be made substantially more accurate by augmenting the estimator state with foot velocities and modeling the stance foot's motion as rolling (v_f = ω_f × r) instead of stationary, and by wrapping this rolling-aware error-state Kalman filter in an interacting multiple-model (IMM) filter with two contact modes (rolling and slip) that differ only in foot-velocity process noise. In extensive simulations and real-world experiments on a quadruped, this design reduces ATE position error by roughly half to two-thirds relative to state-of-the-art point-contact filters and single-model rolling filters, on flat, slippery, uneven, a","pith_inferences":["The paper attributes the gains to the rolling model, but the ablation does not test a foot-velocity state with the standard zero-velocity constraint. If the gains are largely due to the extra foot-velocity state acting as a noise-shaping buffer rather than the rolling constraint itself, the mechanism would need to be reinterpreted, although the empirical results could still hold.","On terrain where foot deformation is significant (e.g., soft ground), the rolling model's assumption of a single radius vector r will break down; a promising extension is to treat r as an uncertain parameter or to add a third mode for deformed contact.","The IMM mode probabilities could serve as an online slip/contact-quality signal, which could be fed back to the locomotion controller to reduce slip proactively.","The analysis of the learned baseline suggests that implicit recurrent adaptation struggles with long-horizon contact switching; combining IMM's explicit mode structure with learned innovation models might yield robust hybrid estimators, though the paper does not explore this."],"forward_implications":["If correct, legged robots can operate with significantly lower drift in GPS-denied or vision-degraded environments, relying only on IMU and joint encoders.","The two-mode IMM structure can be extended to more contact modes (finer noise discretization) or combined with exteroceptive measurements when available, potentially improving full state-estimation pipelines.","Explicit foot-velocity estimation offers a natural slip/contact-quality signal that could be fed to gait control or planning.","The rolling contact model is simple and frame-agnostic, so it may transfer to other legged platforms with spherical or curved feet without retraining.","The computational overhead of IMM is modest (about 1 ms per step), making the approach suitable for real-time embedded deployment."],"fun_headline_variants":["Rolling-foot model cuts legged-robot drift by half","IMM filter halves pose error in legged odometry","Rolling contact filter beats point-contact odometry","Two-mode foot model slashes legged-robot drift","IMM proprioceptive odometry improves legged pose"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The filter assumes that a stance foot's true velocity is exactly the kinematic rolling velocity ω_f × r, with any deviation attributable to noise; if actual foot motion on a given terrain is not rolling-like (e.g., due to deformation, soft ground, or highly uneven surfaces), the rolling-aware improvement may vanish.","fun_headline_variants_meta":{"raw":{"variants":["Rolling-foot model cuts legged-robot drift by half","IMM filter halves pose error in legged odometry","Rolling contact filter beats point-contact odometry","Two-mode foot model slashes legged-robot drift","IMM proprioceptive odometry improves legged pose"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000448,"raw_usage":{"total_tokens":2069,"prompt_tokens":686,"completion_tokens":1383,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":1304}},"tokens_in":430,"tokens_out":1383,"duration_ms":10459,"temperature":1.0,"reasoning_tokens":1304,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:03:57.176733+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare IMM-PO against a single-model point-contact IEKF on a soft or highly uneven terrain where rolling is not a good model; if the point-contact filter matches or beats IMM-PO, the rolling premise is not the source of the reported gains. Alternatively, remove the third measurement block in Eq. 16 (the rolling constraint on foot velocity) and keep the foot-velocity state; if accuracy is unchanged, the gains come from the state augmentation, not the rolling model.","supporting_citations":[],"review_version":1}