{"id":"0ac28a01-9f39-47d3-91ba-0ede1fc012d1","arxiv_id":"2506.19968","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A training-free pipeline that selects a stable gait and optimizes leg tip trajectories with differential evolution restored forward walking in a damaged hexapod.","lead":"The paper presents a two-stage algorithm that lets a six-legged robot with damaged legs quickly learn a new walking pattern without any pre-training. The method was tested on a real hexapod and restored forward walking after one or two legs failed.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2 as printed schedules the same leg to swing twice and leaves another functional leg in permanent support in every single-leg-loss case, contradicting the paper's stated rule-of-neighborhood stability design; it is therefore unclear whether the hardware validation used the algorithm as…","rationale":"The central claim, that rule-based gait switching plus DE over 14 trajectory parameters restores forward locomotion without training or trial-and-error, is partially supported: Table II reports forward progress after recovery in all four hardware scenarios, and the paper is explicit about its assumptions (morphology vector known, legs detached) and its admitted sim-to-real deviations. I therefore do not challenge the conditional verdict. I do disagree with the reader about where the weakest link sits. The reader's sim-to-real concern is real but bounded: the paper documents the gap in Section IV-C, and the headline claim (restored locomotion) survives it. The sharper problem is an internal inconsistency in the single-leg-loss planner. Algorithm 2's λ bookkeeping makes the third phase re-swing β2 and leaves a functional leg permanently in support in every one-leg-loss case, so the gait the pseudocode generates is not the modified quadrangular gait the paper describes and is not covered by the rule-of-neighborhood stability argument cited from [32]. Because the planner is fully specified, this is checkable purely from the text; it is not a disagreement with the field's consensus. The consequence is that the hardware validation may not validate the algorithm as published, and if it did validate the literal algorithm, the permanent-support leg drags, connecting the admitted friction-model limitation directly to the gait design. Secondary weaknesses I did not elevate: no wall-clock timing supports the 'within one hour' headline (1800 evaluations of 10-second simulations at an unreported speed-up), and Table II averages 10 runs without error bars. These are consistent with the conditional verdict. In fairness, the DE configuration (Table I), the closed-form IK in Eq. (11), and the explicit trajectories in Eq. (3) are re-derivable, and the fault, if confirmed, appears localized to Algorithm 2's phase-3 assignment; a correction plus re-validation would likely suffice rather than a rejection of the empirical findings.","tokens_in":10937,"tokens_out":29435,"duration_ms":270045,"concrete_test":"Run Algorithm 2 exactly as printed for X=[0,1,1,1,1,1] (leg 1 missing): if the phase-1 and phase-3 swing groups both contain leg 3 (i.e., β2) while leg 6 is never scheduled to swing, the planner contradicts the Section III-A design. Then re-run the 'Leg 1 missing' scenario of Section IV in the [27] simulator with this literal schedule and compare the CoM and yaw traces with Figures 10-11: matching traces mean the validated gait violates the stated rule of neighborhood, while mismatching traces mean the published pseudocode differs from the tested implementation. Either outcome forces a correction to Algorithm 2 or to the stability claims, and settles the reproducibility question.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Executing Algorithm 2 literally for any one-leg-loss morphology reveals an internal inconsistency with the stability design claimed in Section III-A. The 'if γ=3 then λ=i' test fires on the second functional leg of the group being scanned, and because the second loop over L2={2,3,6} always runs last, it unconditionally overwrites λ, giving λ≡β2 in all six single-leg cases. The printed schedule then has leg set {α1,β2} swing in phase 1, {α2,β1} in phase 2, and {λ=β2} again in phase 3: the same leg swings twice per cycle, and the third functional leg of the group with three members (β3 when the lost leg is in L1, α3 when it is in L2) never enters swing. Example, with leg 1 missing: swing sets {4,3}, {5,2}, {3}, and leg 6 never lifts. This contradicts (a) the claimed incorporation of the rule of neighborhood, whose analytic stability guarantee in [32] therefore does not apply to the generated gait; (b) the description of the modified quadrangular gait as alternating groups of two legs; and (c) recovery behavior that uses all remaining legs. The load-bearing consequence is a reproducibility gap: the hardware experiments in Section IV either validated a gait different from the one Algorithm 2 produces, or validated a gait that violates the paper's own stability rationale. In the latter case the permanent-support leg is dragged and slides every cycle, which is precisely the regime where the paper's admitted simplified friction model is least reliable; the reported sim-to-real yaw gap may then be a downstream symptom of this gait error rather than an independent limitation. Either way, the paper as written does not establish that the algorithm it presents is the algorithm it validated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free damage recovery pipeline for multi-legged robots. After a leg is lost, the method first selects a new gait sequence from a small set of canonical patterns (modified quadrangular for one missing leg, modified pentagonal for two missing legs), and then optimizes 14 leg-trajectory parameters with a differential evolution algorithm whose cost function is evaluated in a closed-form whole-body dynamics simulator developed in the authors' prior work. The approach is validated on a 24-DoF Hiwonder JetHexa hexapod in four damage scenarios involving one or two missing legs; the reported results show that the robot reacquires forward locomotion after optimization, and the abstract claims that recovery is achieved within one hour without pre-training or physical trial-and-error.","tokens_in":1808,"tokens_out":2472,"duration_ms":94063,"significance":"The real-hardware validation is a genuine strength, and the idea of decomposing damage recovery into rule-based gait sequence selection followed by trajectory optimization is practically appealing. If the method works exactly as described, the contribution is a simple, inexpensive alternative to learning-based or trial-and-error approaches, and the claimed ability to handle multiple leg losses goes beyond several prior single-failure methods. The paper also benefits from being grounded in the authors' previously published fast dynamic simulator, which gives the optimization loop a concrete computational basis. However, the as-printed gait sequence algorithm is internally inconsistent with the stated stability rationale and with the trajectory model used in the optimizer, and several headline quantitative claims are not backed by reported measurements. These issues prevent the paper from being accepted in its current form.","major_comments":[{"comment":"As written, Algorithm 2 always sets lambda to beta_2, the second functional leg encountered while scanning L2, regardless of which leg is damaged. The condition 'if gamma = 3 then lambda = i' fires on the second functional leg of a group, and because the loop over L2 runs last, it unconditionally overwrites any value set in the L1 loop. For example, with leg 1 missing, the printed swing sets are {4,3}, {5,2}, {3}: leg 3 swings in both phase 1 and phase 3, and leg 6 never leaves the support phase. This contradicts (a) the claimed incorporation of the rule of neighborhood from [32], since leg 3 swings in the same phase as its neighbor leg 4; (b) the description of the modified quadrangular gait as alternating groups of two legs; and (c) the trajectory model in Eq. (3), in which each leg has exactly one swing phase per gait cycle. The hardware experiments in Section IV therefore either validated a gait different from the one that Algorithm 2 produces, or validated a gait that violates the paper's own stability design. This is a load-bearing reproducibility gap in the central claim.","section":"Section III-A, Algorithm 2"},{"comment":"The statement that locomotion is 'reliably restored in under one hour' is not supported by any timing measurement. The paper reports that the DE objective converges after about 30 generations, but it does not report the wall-clock time per simulator evaluation, the total optimization time, whether evaluations were parallelized, or the time required for gait sequence selection and deployment. With Ps=30, 30 generations imply about 900 simulator evaluations, but no runtime data are given. The one-hour claim should either be removed or backed by an actual end-to-end time measurement.","section":"Abstract and Section IV-B"},{"comment":"The paper states that each damage scenario is run 10 times to demonstrate robustness, but Table II reports only average values with no standard deviations, confidence intervals, or individual run data. The convergence figures show only the best objective value per generation, without clarifying whether these are single runs, averaged curves, or all runs overlaid. Moreover, the table appears to average the 10 best simulated solutions rather than repeated physical experiments; it is unclear how many hardware trials were performed per scenario. Without variance information, the claimed robustness and convergence of the algorithm are not statistically substantiated.","section":"Section IV-C and Table II"},{"comment":"The entire efficiency argument rests on selecting the best simulated candidate and deploying it on the robot with no trial-and-error, but the paper concedes that real-world translations and yaw deviate from simulation due to simplified friction, unmodeled dynamics, and joint saturation. The figures show simulated and real traces, yet the paper provides no quantitative sim-to-real error metrics, such as RMSE in position or yaw, and no evidence that the simulator ranks candidate gaits in the same order as the physical robot. Without this evidence, the reader cannot assess whether the optimized gait is near-optimal on hardware or merely a feasible one, which weakens the 'optimal gait reconfiguration' and 'avoids trial-and-error' contributions.","section":"Section IV-C, simulator-to-real transfer"}],"minor_comments":[{"comment":"The caption says 'Leg 3 is missing', while Table II and the text describe the fourth scenario as 'Leg 4 missing'. Please correct this inconsistency, and unify the 'missed'/'missing' wording throughout Table II and the text.","section":"Figure 12 caption"},{"comment":"The support-phase expression for y appears to have a mismatched parenthesis, and the notation Ng/sigma should clarify whether integer division or floor is intended. The phase indices k = 1,...,Ng/sigma imply that Ng is divisible by sigma; if not, the gait timing is ambiguous.","section":"Section III-B, Eq. (3)"},{"comment":"The printed Xmax and Xmin vectors have no index mapping to the variables in Eq. (4). Please state explicitly which entries correspond to y01,...,y06, x01,...,x06, Lsl, and Lsh, and specify exactly how the 20% relaxation for legs adjacent to a damaged leg is applied to these bounds.","section":"Section IV-B, bound vectors"},{"comment":"References [22] and [29] are the same paper (Erden and Leblebicioğlu, 'Free gait generation with reinforcement learning for a six-legged robot', Robotics and Autonomous Systems, 2008). Please remove the duplicate.","section":"References"},{"comment":"There is a typo, 'optimizatoin', in the sentence introducing the optimization problem. Also, the objective weights w1,...,w5 are introduced without any sensitivity analysis or justification beyond a reference to [25]; a short paragraph on how the weights were chosen would improve reproducibility.","section":"Section III-B"},{"comment":"The convergence criterion is not defined; the statement that the objective 'consistently converged within 30 generations' appears to be based on visual inspection of the figures. Reporting the specific termination rule (e.g., a tolerance on the objective change or a fixed generation count) would make the claim precise.","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"The Algorithm 2 issue is the most serious concern. I recommend asking the authors to provide the exact gait sequence used in the hardware experiments, ideally as a table of swing sets for each damage scenario, and to clarify whether the experiments used a corrected version of Algorithm 2. I would also request the missing timing data and per-run statistics before this paper is accepted. The duplicate of references [22] and [29] suggests the reference list was not carefully checked; the authors should audit the bibliography. The mathematical framework and the real-robot demonstrations are promising, but the internal inconsistency between the printed algorithm and the experimental validation must be resolved first."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper has a real hardware validation: four leg-loss scenarios on a 24-DoF hexapod, training-free recovery using a two-stage pipeline (rule-based gait selection plus DE over 14 trajectory parameters), with results from a motion tracker, not just simulation. That is genuinely useful for the field. Second, the printed Algorithm 2 does not match the paper's own description. For any one-leg-loss case, the condition 'γ=3' fires on the second functional leg in each group, and the second loop overwrites λ, so the same leg swings in both phase 1 and phase 3, and one functional leg never leaves support. Example: with leg 1 missing, swing sets are {4,3}, {5,2}, {3}, and leg 6 is permanently in support. That contradicts the claimed rule of neighborhood, the 'modified quadrangular gait' description, and the stability rationale from [32]. The hardware experiments either used a different gait than the one printed, or validated a gait that drags a leg every cycle. Either way, the paper as written does not establish that the algorithm it presents is the algorithm it validated.\n\nWhat's new and earned: the two-stage decomposition—stabilize first via a hand-designed gait sequence, then optimize leg trajectories with DE—is a sensible simplification compared to end-to-end learning or trial-and-error. The simulator from [27],[28] is faster-than-real-time and the physical robot is an independent benchmark, so the circularity burden is low. The objective function in (5) is reasonable, and the convergence plots show consistent behavior across 10 runs.\n\nSoft spots beyond the Algorithm 2 issue: the 'under one hour' claim is not directly evidenced. They report convergence in about 30 generations but no wall-clock time; with a 30-population, 60-generation DE and 10-second simulations, an hour would require parallelism or speedups that are not documented. Also no error bars or variance for the reported averages in Table II, so 'robustness' is not quantified. The sim-to-real yaw gap is admitted but not explained; the permanent-support leg might be the cause, which would make the gap a symptom of the gait bug rather than an independent modeling limitation.\n\nWho this is for: legged locomotion and fault-tolerance researchers. They get a useful data point and a clear pipeline description. It deserves a serious referee—the hardware effort is real and the bug may be fixable—but it should not be accepted without the authors reproducing Algorithm 2's behavior or correcting it, and reporting actual timing and variance. My recommendation: engage, but treat the presented gait sequence as suspect until the authors respond.","headline":"Real hardware validation of a two-stage evolutionary gait recovery pipeline, but the printed gait sequence algorithm has a load-bearing bug that breaks the paper's own stability rationale.","tokens_in":11887,"tokens_out":4692,"would_cite":false,"duration_ms":44984,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hexapod robot that loses one or two legs can regain forward walking within an hour by switching to a statically stable gait and tuning 14 leg-trajectory parameters with differential evolution, entirely offline.","keywords":["multi-legged robot","damage recovery","gait reconfiguration","differential evolution","hexapod","trajectory optimization","self-modeling","locomotion"],"falsifier":"Run the described algorithm on a 24-DoF hexapod with one leg removed; if the best offline-evaluated parameter set fails to keep the robot walking for a ten-second trial, or if the realized yaw exceeds the roughly 20-degree rotation observed in the paper's hardware tests, the transfer claim is contradicted.","tokens_in":10734,"feed_emoji":"🦿","tokens_out":9030,"duration_ms":80539,"temperature":0.7,"pith_summary":"This paper claims that a multi-legged robot that loses one or two legs can recover forward locomotion in under an hour, without any pre-training or physical trial-and-error. Recovery is decomposed into two stages: a rule-based switch to a stable canonical gait appropriate for the number of remaining functional legs, and a differential-evolution search over 14 leg-trajectory parameters whose fitness is evaluated in a faster-than-real-time simulator. The work tests this on a 24-degree-of-freedom hexapod across four damage scenarios, including loss of one leg and loss of two legs; in the most severe cases the robot could not stand before recovery and walked after it. A sympathetic reader would care because the method promises cheap, onboard-compatible damage adaptation for field robots.","feed_headline":"Hexapod relearns to walk an hour after leg loss","feed_subtitle":"A rule-based gait switch plus 14-parameter evolution restores forward motion with no pre-training.","key_machinery":"The load-bearing machinery is the pair formed by the morphology vector $X$ and the faster-than-real-time simulator developed in [27], [28]: $X$ tells the planner which legs remain and therefore which canonical gait satisfies the three-support-leg stability condition, while the modular, closed-form dynamic model, written in a body-coordinate Lagrangian formulation, lets the differential-evolution optimizer evaluate each candidate entirely offline, eliminating trial-and-error on the damaged robot. The optimizer uses classic differential evolution [33] with a current-to-best mutation and binomial crossover, operating on the 14-dimensional trajectory parameter vector, with inverse kinematics from [35] converting tip trajectories into joint commands and a joint-space PID controller tracking them. The objective function in (5) is the scoring rule that couples all of these pieces.","core_discovery":"The paper's central claim is that leg damage in a multi-legged robot can be encoded as a binary morphology vector $X$, and that this vector alone is enough to select both a new gait sequence and a feasible search space for trajectory reconfiguration. For a hexapod, the healthy tripod gait is replaced by a modified quadrangular gait when one leg is lost and by a modified pentagonal gait when two are lost; the rule of neighborhood guarantees static stability. Differential evolution then optimizes the per-leg trajectory offsets $x_{0i}$, $y_{0i}$ together with global step length $L_{sl}$ and step height $L_{sh}$, scored by $$F = \\frac{w_1 $y_f^{2}$}{1 + w_2 $x_f^{2}$ + w_3 \\$psi_f^{2}$ + w_4 \\$\\Delta$\\$phi^{2}$ + w_5 \\$\\Delta$\\$theta^{2}$},$$ which rewards forward travel and punishes lateral drift, yaw, and roll/pitch oscillation. In all tested scenarios the objective converged within about 30 generations, and the best simulated parameter set restored straight walking on the physical robot with lateral drift and yaw kept within roughly 20 degrees.","pith_inferences":["Editorial: the two-stage decomposition suggests a general template—rule-based gait selection for the discrete structure, continuous optimization for the trajectory parameters—that could transfer to quadrupeds or eight-legged robots, but the paper only demonstrates hexapods.","Editorial: the reported simulation-to-reality gaps in translation and yaw, attributed to simplified friction and joint saturation, imply that a closed-loop correction step after transfer might close the gap; the paper leaves this to future work.","Editorial: the binary morphology vector treats a damaged leg as fully detached; extending the same scheme to locked joints or partial limb loss would require enriching $X$ and revisiting the stability rule.","Editorial: because only the best solution is transferred and differential evolution is stochastic, repeated runs yield slightly different gaits; a testable extension would compare the distribution of realized trajectories across repeated runs."],"forward_implications":["A hexapod with five intact legs switches to a modified quadrangular gait; one with four intact legs switches to a modified pentagonal gait; both are statically stable under the rule of neighborhood.","Convergence within roughly 30 generations with population size 30 implies the full recovery search fits inside one hour even on the robot's onboard class of computer.","Because every candidate is evaluated in simulation, the approach avoids further wear or damage from testing unstable candidate gaits on the injured robot.","The same objective function with the same weights serves all four damage scenarios, so no per-scenario tuning of the recovery objective is needed.","Recovery requires at least two functional legs per side; with fewer, the planner returns an error rather than a gait."],"supporting_citations":[{"why":"supplies the modular faster-than-real-time dynamic model used to evaluate every candidate in the offline optimization","marker":"[27]"},{"why":"provides the singularity-free whole-body equations used when simulating damaged morphologies","marker":"[28]"},{"why":"introduces the differential evolution algorithm that drives the trajectory search","marker":"[33]"},{"why":"gives the closed-form inverse kinematics that converts optimized tip trajectories into joint-angle commands for the 3-DoF legs","marker":"[35]"},{"why":"supplies the earlier objective-function study that the paper's fitness function builds on","marker":"[25]"},{"why":"is cited for the rule of neighborhood that guarantees statically stable gait sequences","marker":"[29]"},{"why":"provides the analytical proof that gaits satisfying the rule of neighborhood are statically stable","marker":"[32]"}],"fun_headline_variants":["Damaged hexapod walks again in one hour, no training","Hexapod relearns walking an hour after leg loss, no pre-training","Evolutionary algorithm fixes hexapod gaits in 60 minutes","Leg injury? Hexapod evolves new walk in an hour flat","Damaged legged robot reconfigures gait without training in 1h"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire recovery depends on the faster-than-real-time simulator being faithful enough that the best simulated gait transfers to the physical robot without trial-and-error; the paper's own hardware tests show real-world translation and yaw deviate from simulation, so that fidelity is only partly demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["Damaged hexapod walks again in one hour, no training","Hexapod relearns walking an hour after leg loss, no pre-training","Evolutionary algorithm fixes hexapod gaits in 60 minutes","Leg injury? Hexapod evolves new walk in an hour flat","Damaged legged robot reconfigures gait without training in 1h"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000499,"raw_usage":{"total_tokens":2407,"prompt_tokens":875,"completion_tokens":1532,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":1437}},"tokens_in":491,"tokens_out":1532,"duration_ms":11722,"temperature":1.0,"reasoning_tokens":1437,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:22:21.323071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the described algorithm on a 24-DoF hexapod with one leg removed; if the best offline-evaluated parameter set fails to keep the robot walking for a ten-second trial, or if the realized yaw exceeds the roughly 20-degree rotation observed in the paper's hardware tests, the transfer claim is contradicted.","supporting_citations":[{"cited_title":"Hexapod robot kinematics modeling and tripod gait design based on the foot end trajectory,","cited_arxiv_id":null,"evidence_quote":"gives the closed-form inverse kinematics that converts optimized tip trajectories into joint-angle commands for the 3-DoF legs"},{"cited_title":"Free gait generation with rein- forcement learning for a six-legged robot,","cited_arxiv_id":null,"evidence_quote":"is cited for the rule of neighborhood that guarantees statically stable gait sequences"},{"cited_title":"Analysis of wave gaits for energy efficiency,","cited_arxiv_id":null,"evidence_quote":"provides the analytical proof that gaits satisfying the rule of neighborhood are statically stable"}],"review_version":2}