{"id":"25bc96e2-2c95-475d-ba54-d98e4505a008","arxiv_id":"2602.17393","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A contact-anchored estimator using footfall records, height clustering, and an inverse-kinematics cubature Kalman filter achieves low-drift proprioceptive odometry on quadruped and wheel-legged robots.","lead":"This paper describes a proprioceptive odometry system that uses foot contacts as world-frame anchors to estimate a legged robot's motion using only inertial and motor data. It reports closed-loop position errors on four quadruped platforms, including wheel-legged robots, and releases code and some datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Wheel-legged odometry relies on no-slip/flat-ground contact propagation (Eqs. 21-25); the paper's own 7.68 m error on the 700 m Astrall C loop is attributed to unmodeled slip, so the central claim of accurate long-horizon wheel-legged odometry is not established.","rationale":"The reader's weakest assumption identified unmodeled wheel slip and flat-ground propagation as the critical weakness. I agree: this is the most load-bearing concern because it directly invalidates the wheel-legged portion of the central claim, and the paper's own experimental evidence (7.68 m error on Astrall C, attributed to slip) demonstrates the failure mode in the exact evaluation used to support the claim. The flat-ground assumption compounds the issue on slopes, which are common in real-world deployments. The point-foot results (Astrall A, B) are impressive and support the claim for point-foot robots on benign terrain, but the title and conclusion encompass wheel-legged robots. The paper explicitly lists slip and slope as limitations, so this is not an external objection but an internal admission. A concrete simulation test can settle whether slip is indeed the dominant error source and whether the method degrades gracefully or catastrophically. The reader's CONDITIONAL verdict remains appropriate: the method is promising for point-foot robots, but the wheel-legged claim requires either slip compensation or a restricted scope statement. I do not recommend changing the verdict; I recommend the authors run the proposed test and either add slip awareness or narrow the claim to flat, high-friction surfaces.","tokens_in":16146,"tokens_out":11158,"duration_ms":108066,"concrete_test":"Run the released Gazebo simulation with the wheel-legged model over a 200 m closed loop under two conditions: (1) high-friction ground, and (2) identical path with a low-friction patch (μ≈0.2) covering 10% of the distance. Compare planar closure errors and inspect the contact-point propagation (Eq. 23) at the patch. If the low-friction run's error exceeds the baseline by more than 2×, unmodeled slip is the dominant error source and the fixed-anchor model fails in the regime the paper claims. Optionally, inject artificial wheel encoder slip in simulation to calibrate sensitivity of error growth to slip magnitude.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section VII) states that contact anchoring provides accurate long-horizon proprioceptive odometry for legged and wheel-legged robots. For wheel-legged platforms, the contact model assumes no wheel–ground slip and locally flat ground: Eq. 21 computes the effective rolling increment by subtracting shank pitch from the wheel encoder, Eq. 23 integrates this into the footfall record, and Eq. 22 projects the rolling direction onto the horizontal plane. Under slip, the encoder increment does not equal the true ground displacement, and because the footfall record is updated with the erroneous rolling displacement, the slip error is baked into the world-frame anchor and accumulates without correction. The paper explicitly acknowledges that slip is not detected or compensated and attributes the 7.68 m error on the 700 m Astrall C loop to slip (Section VI.C-a). Similarly, the flat-ground assumption is listed as a limitation (Section VI.C-c). Thus, the very experiment intended to demonstrate long-horizon wheel-legged capability contains a known violation of the core model, and the reported error is attributed to that violation. Without slip handling or slope-aware propagation, the claim that accurate long-horizon wheel-legged odometry is achieved 'in practice' is not supported beyond flat, high-friction surfaces. The point-foot results are strong, but the wheel-legged claim is central to the paper's scope.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents CAPO, a proprioceptive state estimator for legged and wheel-legged robots that uses only IMU and motor measurements. The method records footfall positions at touchdown, treats stance contacts as world-frame kinematic anchors, adds a support-plane height clustering correction to reduce elevation drift, and propagates wheel contacts by an effective rolling angle. An optional inverse-kinematics cubature Kalman filter suppresses encoder-quantization velocity spikes, and a multi-contact geometric consistency term provides a yaw correction. The estimator is evaluated in Gazebo on an AlienGo model and on four real platforms—Unitree Go2 EDU and three Astrall robots—with closed-loop trajectories. The paper reports, for example, 0.1638 m planar closure error over a ~200 m loop on a point-foot robot, and 7.68 m over a ~700 m loop on a wheel-legged robot. The central claim is that contact anchoring provides accurate long-horizon proprioceptive odometry for both legged and wheel-legged robots.","tokens_in":16613,"tokens_out":4726,"duration_ms":51216,"significance":"If the point-foot results are representative, the paper makes a useful contribution: purely proprioceptive odometry with decimeter-level loop closure over hundreds of meters is notable, and the public release of code and ROS bags is a real strength. The closed-form rolling-bias analysis (Eq. 18) and the CKF recursion are clearly presented and appear internally consistent. However, the paper's evidence for the wheel-legged portion of the central claim is substantially weaker, and the single-run nature of the experimental validation limits confidence. The manuscript would be strengthened by repeated trials, explicit parameter values or sensitivity analysis, and either slip handling for wheel-legged contacts or a narrowed claim.","major_comments":[{"comment":"The wheel-legged claim is not supported by the current evidence. The contact-propagation model assumes no wheel–ground slip and locally flat ground: Eq. (23) integrates the encoder-derived rolling increment into the world-frame footfall record, so under slip the erroneous displacement is baked into the anchor. The paper explicitly acknowledges in Section VI.C-a that slip is not detected or compensated, and attributes the 7.68 m closure error over the 700 m Astrall C loop to slip. That is the only long-horizon wheel-legged experiment, and it contains a known violation of the core model. The 200 m wheel-legged result (0.2264 m) may be on favorable, high-friction flat ground, but one run does not establish long-horizon capability. The Section VII conclusion that contact anchoring provides accurate long-horizon proprioceptive odometry for wheel-legged robots overreaches; either slip-aware pr","section":"Section III-B, Eqs. (21)-(25); Section VI.C-a; Table II"},{"comment":"The real-world validation consists of a single run per condition with no repeated trials, error bars, or statistical summary. Table II reports one closure error per platform/loop, and Section VI.B describes 'representative' traces. With several configurable thresholds and gains, it is impossible to tell whether the reported numbers are typical or selected favorable outcomes. At minimum, the authors should report the number of runs and mean/standard deviation for each condition, or clearly state that these are single demonstrations. This is load-bearing for the claim of accurate 'in practice' odometry because the headline numbers—especially the difference between Astrall A/B and Astrall C—are otherwise anecdotal.","section":"Table II and Section VI.B"},{"comment":"At least six algorithm parameters are configurable but no values are reported: f_th in Eq. (1), Delta_h and T_fade in Eqs. (10)-(11), kappa in Eq. (11), alpha_0 and T_psi in Eq. (35), plus the CKF noise covariances Q and R after Eq. (48). While the code is public, the paper does not state how these were chosen, whether they are fixed across platforms, or how sensitive the results are to them. This matters because a method with many free parameters can be accidentally tuned to the evaluated trajectories. Please provide the exact values used for each platform and a sensitivity or ablation study on at least the most influential thresholds (e.g., Delta_h and f_th).","section":"Eqs. (1), (10), (11), (35); Sec IV-B (Q, R)"}],"minor_comments":[{"comment":"The title says 'Quadruped Robots' but the abstract and introduction claim a unified formulation for biped, quadruped, and wheel-legged robots. Please align the title and scope statements.","section":"Title/Abstract"},{"comment":"The text refers to 'Robot A (MP)' and 'Robot B (MW)' in Fig. 18 without defining MP/MW. Define these abbreviations or use the point-foot/wheel-legged terminology consistently.","section":"Section VI.B-2-a"},{"comment":"In Eq. (18), the notation 'a_1 measured from the +x axis' should be stated in the figure caption as well. The closed-form result is helpful, but the derivation would benefit from a one-line sign convention reminder near the equation.","section":"Section III-A, Fig. 5"},{"comment":"The support-plane height correction is described only in prose. A pseudo-code block or explicit algorithm box for the match/snap/prune logic in Eqs. (9)-(11) would improve reproducibility.","section":"Section II-D"},{"comment":"The simulation comparison is against a LiDAR SLAM baseline. This is reasonable, but the paper does not state whether the SLAM baseline had access to the same closed-loop control or whether it was optimized per trajectory. Please clarify.","section":"Section VI.A"},{"comment":"The yaw correction gain schedule uses alpha_0 and T_psi, but the paper does not state how 'full support' is determined beyond all four feet being in stance. A precise condition (e.g., continuous stance duration) would avoid ambiguity.","section":"Section V-B, Eq. (35)"}],"recommendation":"major_revision","confidential_remarks":"The point-foot results are genuinely interesting, and the public implementation is a clear plus. My main concern is that the paper's central conclusion is broader than the evidence: the wheel-legged 700 m experiment explicitly violates the model's assumptions and is attributed to the very phenomenon the model does not handle. This is fixable by adding slip-aware propagation or narrowing the claim to point-foot legged robots and presenting wheel-legged as preliminary. I would also like to see repeated trials or error bars before the quantitative claims are taken at face value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The point-foot result is the strongest thing here: 0.16 m over a 200 m loop on the Astrall A is genuinely good for purely proprioceptive odometry, and the vertical-loop numbers are also credible. The paper does deserve a serious read.\n\nWhat's actually new is the package as a whole, not any single invention. Contact-aided odometry and kinematic anchors go back to Hartley and others, but this paper ties them together with a few tricks that matter in practice: the height-clustering/time-decay correction that snaps new footfalls to known support planes, the IKVel-CKF to tame encoder-derived velocity spikes, and the multi-contact geometric yaw prior that arrests yaw drift during standing. The height clustering is a particularly nice touch because vertical drift is a known killer in this setting. The authors also ship the code and Go2 ROS bags, which is the right kind of evidence.\n\nWhere I land on the soft spots: the wheel-legged story is not backed up by the data as strongly as the conclusion claims. The 700 m Astrall C loop gives 7.68 m of closure error, and the paper itself attributes that to unmodeled slip. Given that the wheel-contact propagation assumes no slip and locally flat ground (Eqs. 21-23), the central claim of accurate long-horizon wheel-legged odometry is only supported on flat, high-friction surfaces. That is a real limitation, not a nitpick, and it is the one part of the stress-test note I fully endorse. The paper lists flat-ground as a limitation but still closes with the broad claim; the conclusion should be softened or slip handling added.\n\nAlso worth flagging: real-world validation is single-run, no error bars, and at least six parameters (f_th, Delta_h, T_fade, kappa, alpha_0, T_psi) are configurable with values not given in the text. Public code mitigates that, but a sensitivity analysis or at least a table of chosen values and ablation would make the results much easier to trust. The biped claim is not supported by any biped experiment; the video shows a quadruped switching support modes, which is not the same thing.\n\nNone of this kills the paper. The point-foot odometry and the height-correction mechanism are solid enough to warrant referee time, and the code/data release makes reproducibility possible. I would send it to review, but the authors should be pushed to add repeated trials, specify the parameters, and either fix or reframe the wheel-legged claim.","headline":"The point-foot results are the real contribution; the wheel-legged claims are oversold given the paper's own slip caveats.","tokens_in":17038,"tokens_out":1847,"would_cite":true,"duration_ms":19667,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A legged robot can navigate hundreds of meters with only its own joints and an IMU, by locking onto footfall anchors.","keywords":["proprioceptive odometry","legged robots","wheel-legged robots","contact anchoring","footfall records","cubature Kalman filter","dead reckoning","state estimation"],"falsifier":"Run a wheel-legged robot over a known distance on a deliberately slippery surface, such as a wet vinyl floor or loose gravel, while measuring ground truth with motion capture; if the estimated path error grows with the number of slip events and exceeds the paper's reported drift, the no-slip contact-propagation assumption is falsified.","tokens_in":16062,"feed_emoji":"🤖","tokens_out":4027,"duration_ms":36864,"temperature":0.7,"pith_summary":"This paper claims that a legged or wheel-legged robot can estimate its own position accurately over long distances using only proprioceptive sensors—an IMU and joint motor encoders—with no cameras or LiDAR. The key idea is to treat each footfall as a kinematic anchor: at touchdown, the foot's world position is recorded and treated as a fixed constraint while the foot stays on the ground. These intermittent anchors suppress the drift that normally accumulates in inertial integration. In closed-loop experiments, a point-foot quadruped returns to its start with just 0.16 m error after a roughly 200 m walk, and a wheel-legged robot manages 0.23 m on the same course. The paper also introduces mechanisms to control vertical drift (height clustering of touchdown heights) and yaw drift (geometric consistency of multiple contacts).","feed_headline":"Footfalls as landmarks rein in drift for leg-only odometry","feed_subtitle":"Closed-loop tests on quadrupeds show a point-foot robot closing a 200 m loop with 0.16 m error, using no cameras or LiDAR.","key_machinery":"The central object is the footfall record: the world-frame position of an end-effector recorded at touchdown and assumed stationary during stance. During stance, the robot's body pose is inferred by subtracting the current forward-kinematics end-effector position (rotated into the world frame) from the stored footfall point. This converts each stance phase into an intermittent absolute position constraint, suppressing long-horizon drift without exteroceptive sensing. Surrounding mechanisms include torque-based wrench stance detection, support-plane height clustering with time decay, effective wheel rolling compensation for wheel-legged platforms, an inverse-kinematics cubature Kalman filter","core_discovery":"The central claim is that contact anchoring with touchdown footfall records provides accurate long-horizon proprioceptive odometry in practice. The estimator records the world-frame foot position at each touchdown, treats it as stationary during stance, and derives body position and velocity constraints from forward kinematics. A torque-based wrench estimate selects which legs are in contact. A support-plane height clustering snaps newly recorded footfall heights to previously observed heights, preventing elevation drift. For wheel-legged robots, the recorded contact is propagated by an effective wheel rolling angle that subtracts shank-pitch-induced encoder motion, and a cubature Kalman fil","pith_inferences":["If slip detection and weighting were added to the wheel-legged branch, the large horizontal error on the 700 m wheel-legged loop (7.68 m) might be reduced to near point-foot levels, a testable extension of the paper's own analysis.","The height-clustering step effectively learns a discrete map of support planes from touchdown history; this could be repurposed as a lightweight terrain representation for footstep planning or traversability assessment.","The multi-contact yaw correction could be combined with a magnetometer or with occasional visual heading fixes to provide a redundant heading reference that is robust to both IMU drift and contact-model violations.","The local-flatness assumption in wheel-contact propagation could be relaxed by estimating a sloped support plane from multi-contact geometry, which would broaden the estimator's applicability to ramps and general uneven terrain."],"forward_implications":["Legged robots can maintain accurate pose estimates in low-texture or illumination-challenged environments where camera- or LiDAR-based SLAM fails, using only onboard proprioception.","Elevation drift over long traversals can be controlled by clustering touchdown heights onto previously observed support planes, even over repeated stair ascents and descents.","Wheel-legged platforms can be folded into the same estimator by propagating the contact point through wheel rolling, but the accuracy depends on the no-slip and locally-flat-ground assumptions.","Yaw drift, a known weak point of inertial-only estimation, can be arrested by enforcing geometric consistency of multiple stance contacts, providing a fallback heading reference when IMU yaw is unreliable.","The contact-set formulation is morphology-agnostic and extends to bipeds and multi-limbed robots by varying the number of contacting end-effectors over time."],"fun_headline_variants":["Touchdowns anchor pose: camera-free legged odometry with low drift","Purely proprioceptive odometry: footfalls rein in IMU drift","Legged Odom without cameras: contact anchors beat drift","Footfalls as landmarks cut drift in legged odometry","Wheel-legged and legged robots: footfall landmarks erase drift"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"For wheel-legged robots, the contact propagation assumes no wheel–ground slip and locally flat ground; slip is not detected or compensated, so the long-horizon accuracy claim for those platforms depends on the ground being non-slippery and locally flat.","fun_headline_variants_meta":{"raw":{"variants":["Touchdowns anchor pose: camera-free legged odometry with low drift","Purely proprioceptive odometry: footfalls rein in IMU drift","Legged Odom without cameras: contact anchors beat drift","Footfalls as landmarks cut drift in legged odometry","Wheel-legged and legged robots: footfall landmarks erase drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1219,"prompt_tokens":756,"completion_tokens":463,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":373}},"tokens_in":500,"tokens_out":463,"duration_ms":4856,"temperature":1.0,"reasoning_tokens":373,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T22:13:24.075163+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a wheel-legged robot over a known distance on a deliberately slippery surface, such as a wet vinyl floor or loose gravel, while measuring ground truth with motion capture; if the estimated path error grows with the number of slip events and exceeds the paper's reported drift, the no-slip contact-propagation assumption is falsified.","supporting_citations":[],"review_version":1}