{"id":"60133d0d-d57e-481a-945a-221ee5763d82","arxiv_id":"2412.06776","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A differentiable 'sum of Lyapunov exponents' score is used as a robustness objective to co-optimize robot hardware and control policies in simulation.","lead":"Robot designers usually tune stiffness, damping, and link shapes by trial and error. This paper proposes a mathematical score, built from Lyapunov exponents, to measure how quickly small disturbances grow, and then lets a simulator adjust the robot's design and controller to make that score as negative as possible.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central metric Lλ=1^T λ does not imply 'trajectories not diverging': a negative sum of Lyapunov exponents can coexist with a positive largest exponent, so optimizing Eq. (9) may select chaotic/anisotropic systems that fail the paper's own robustness criterion.","rationale":"I read the paper's central claim as: optimizing the signed sum of Lyapunov exponent components through differentiable simulation is an effective way to improve robustness. For that claim to hold, Lλ must at least order systems correctly on a controlled perturbation-rejection test. The weakest point is indeed the proxy assumption behind Eq. (9), as the reader identified. I agree with that diagnosis but sharpen it from a vague 'overdamped/degenerate designs' concern to a precise mathematical failure mode: the signed sum can be very negative while the largest Lyapunov exponent is positive, so the system is volume-contracting yet still has trajectories that diverge. This is not a disagreement with consensus; it is an internal inconsistency with the paper's own definition of robustness as non-divergence. Algorithm 1's lack of tangent-basis reorthonormalization is a separate correctness issue for the individual spectrum, but the summed quantity is the one actually used in Eq. (9), and the counterexample targets that sum directly. The absence of any quantitative disturbance-amplification metric in Section IV means the empirical demonstrations do not resolve the issue. A single controlled 2D linear test would settle whether more negative Lλ actually implies better finite-perturbation rejection. If it fails, the verdict should move toward rejection or the method should be revised to optimize the largest exponent or a finite-time maximum amplification factor. Since that test is not yet present, I recommend keeping a conditional verdict rather than accepting the central claim as established.","tokens_in":9105,"tokens_out":7867,"duration_ms":92123,"concrete_test":"Use the same Jacobian/rollout machinery on a controlled 2D linear system: variant A=diag(1.05,0.1) and variant B=diag(0.8,0.8). Compute Lλ via Algorithm 1 and via the exact formula log|det A|. Then apply the disturbance protocol analogous to Section IV-B: 100 random initial offsets, recording the maximum state deviation and whether deviations decay over the 2 s horizon. If variant A, despite the more negative Lλ, shows growing deviations while variant B rejects all tested perturbations, Eq. (9) fails as a robustness objective. This directly settles the proxy question without relying on high-DOF contact effects.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section III defines robustness as 'trajectories which are not diverging one from the others' and proposes Eq. (9), Lλ=1^T λ. This identification is not valid: the sum of Lyapunov exponents is the average logarithmic volume contraction rate, while non-divergence of nearby trajectories is governed by the largest exponent, λ_max. Lλ<0 can coexist with λ_max>0, so a system can have globally contracting volume yet exponentially separating trajectories. A simple linear map makes this concrete: A1=diag(1.05,0.1) has exponents (ln 1.05, ln 0.1), giving Lλ≈-2.25, but it is unstable along the first coordinate; A2=diag(0.8,0.8) has Lλ≈-0.45 and contracts every direction. The proposed objective ranks A1 as more robust than A2, contradicting the stated criterion and any finite-perturbation rejection test. Algorithm 1 also computes per-step singular values without reorthonormalizing a tangent basis, so it does not return the Oseledets spectrum; only the summed quantity has a volume-contraction reading, and that summed quantity is exactly the one shown above to be insufficient. The Section IV experiments do not resolve this because no quantitative disturbance-amplification metric or baseline is reported, and the quadruped Lλ change is small (-0.56 to -0.61).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a differentiable robustness metric for robotic systems, defined as the signed sum of Lyapunov exponents Lλ = 1^T λ, and uses it as an optimization objective in a differentiable-simulation framework. The method is demonstrated on three systems: a two-link planar manipulator whose link lengths, masses, and PD gains are co-optimized; a 16-joint spider robot whose joint PD gains are varied to assess falling robustness; and a quadruped whose periodic gait policy parameters are optimized. The central claim is that minimizing Lλ improves the ability of a robot to reject perturbations, with the paper reporting more negative Lλ values and qualitatively more convergent trajectories after optimization.","tokens_in":9369,"tokens_out":7732,"duration_ms":83901,"significance":"The paper addresses an important and open problem: obtaining a quantitative, computable robustness measure for high-degree-of-freedom, contact-rich robotic systems. The computational pipeline, which differentiates through long simulations to optimize a stability-related objective, is creative and potentially valuable for robot co-design. The range of test cases is a strength, and the volume-contraction interpretation of Lλ is mathematically meaningful. However, the central identification of Lλ with 'non-divergence of trajectories' is not valid in general, and the experimental validation is largely circular because the optimized quantity is used as its own evidence of success. With a corrected metric and independent finite-perturbation validation, the approach could be a solid contribution; as presented, the main claim is not yet supported.","major_comments":[{"comment":"The paper defines robustness as 'trajectories which are not diverging one from the others' and selects Lλ = 1^T λ as the metric. This identification is not valid: the sum of Lyapunov exponents measures the average logarithmic rate of phase-space volume contraction, whereas non-divergence of nearby trajectories in every direction is governed by the largest exponent, λ_max. A negative sum can coexist with a positive largest exponent. For example, the linear map A1 = diag(1.05, 0.1) has Lλ = ln(1.05) + ln(0.1) ≈ -2.25 but is unstable along the first coordinate, while A2 = diag(0.8, 0.8) has Lλ ≈ -0.45 and contracts every direction. The proposed objective would rank A1 as more robust, contradicting the stated criterion. The authors should either replace Lλ with a metric that constrains λ_max < 0 or uses the full spectrum, or provide a concrete argument and experimental evidence that the robotic systems considered cannot exhibit this phenomenon.","section":"Section III, Eq. (9)"},{"comment":"The numerical procedure averages the log singular values of each local Jacobian dΦ(x_i) without reorthonormalizing a tangent basis between steps, so it does not compute the Oseledets Lyapunov spectrum. The individual components of Eq. (8) are not the Lyapunov exponents in general; only the sum of the components has a valid volume-contraction reading, because log det(∏ dΦ_i) = ∑ log det(dΦ_i). Consequently, the Van der Pol 'spectrum' in Fig. 2 cannot validate the individual exponents. The paper should either implement a standard QR/SVD algorithm with periodic reorthonormalization (e.g., the Benettin method) or explicitly restrict all claims to the summed quantity Lλ.","section":"Section II-D, Eq. (8) and Algorithm 1"},{"comment":"The experimental validation is insufficient because no independent, non-Lλ robustness measure is reported. The manipulator and quadruped results show that the optimized parameters improve Lλ, but since Lλ was the optimization objective, this is partly by construction; the random-start trajectory plots are qualitative and are not summarized by any quantitative disturbance-amplification or task-success metric. The quadruped change (Lλ from -0.56 to -0.61) is small and is not tied to any behavioral improvement. I recommend adding quantitative comparisons under finite perturbations (e.g., impulse disturbances, actuator noise, terrain variations), reporting quantities such as maximum deviation, settling time, or success rate, and comparing against a baseline optimized for the task alone.","section":"Section IV-B through IV-D"},{"comment":"The theoretical framework assumes an autonomous transition map Φ(x_i), and Section III-A states that u_i = Π(x_i) makes the system autonomous. However, the quadruped policy in Eqs. (11)-(12) is explicitly time-dependent through the sinusoidal reference q_ref(t). A time-periodically forced system is not autonomous, and the standard Oseledets Lyapunov exponents are not directly applicable without augmenting the state with time (or using Floquet theory for periodic orbits). The paper should clarify how Eq. (1) applies to this experiment or use a time-augmented formulation.","section":"Section IV-D and Section III-A"}],"minor_comments":[{"comment":"The sentence 'As in floating base robotics, it may happen that dΦ_i is not a square matrix applying (8) is not viable' appears to refer to Eq. (7), not Eq. (8), and should be restructured for readability.","section":"Section II-D"},{"comment":"The notation |dΦ(x_i)| for a matrix is undefined; the volume expansion/contraction statement should be phrased in terms of the determinant or the singular values of the Jacobian.","section":"Section II-B"},{"comment":"The phrase 'We compare the the results' contains a duplicated article.","section":"Section IV-D"},{"comment":"The word 'Lypaunov exponents' in the conclusions is a typo; it should be 'Lyapunov exponents'.","section":"Section V"},{"comment":"The caption reads 'The two plot on the left' and should read 'The two plots on the left'.","section":"Fig. 6 caption"},{"comment":"The phrase 'where : 1 is the ones vector' contains a stray colon and should be 'where 1 is the ones vector'.","section":"Section III"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the validity of Lλ as a robustness proxy lands: the paper's central derivation equates volume contraction with non-divergence of trajectories, which is not correct. I do not recommend rejection, because the computational pipeline is promising and the volume-contraction interpretation is sound; however, the manuscript needs either a reformulated metric (e.g., one that constrains the largest exponent) or substantial additional validation against independent finite-perturbation benchmarks before the central claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a new application — making the sum of Lyapunov exponents a differentiable objective inside MJX/JAX contact-rich simulation and optimizing hardware/policy parameters against it — but the paper's central metric does not mean what the authors say it means. The sum Lλ is average log volume contraction; their own robustness criterion is \"trajectories which are not diverging one from the others,\" which is governed by the largest exponent. A negative sum can coexist with a positive largest exponent (diag(1.05,0.1) vs diag(0.8,0.8)). So optimizing Eq. (9) can select systems that fail the paper's own definition of robustness. That is a load-bearing flaw, not a nitpick.\n\nCredit where earned: the differentiable formulation is a real practical step. The summed quantity does equal the average logarithmic volume contraction even without tangent-basis reorthonormalization, because log det of the product is the sum of local log determinants. The Van der Pol sanity check is decent, and the random-start trajectory plots for the manipulator suggest the optimized design is more convergent in practice. The paper is honest about gradient-quality limitations in loco-manipulation. Novelty is real: no prior work optimized Lyapunov exponents directly in differentiable contact-rich simulation.\n\nSoft spots, in proportion: (1) the metric choice is not supported by any quantitative link to finite perturbation rejection, actuator noise, or task performance, and no independent non-Lλ robustness benchmark is reported; (2) the individual Lyapunov exponents from Algorithm 1 are not the true Oseledets spectrum, and the only defensible reading is the summed volume-contraction rate, which is exactly the quantity shown to be insufficient; (3) the quadruped policy is time-dependent, so the autonomous-map theory does not directly apply, and the reported Lλ change is small (-0.56 to -0.61) with no error bars or disturbance tests; (4) no code or data are released, making verification hard. The citation pattern looks reasonable, including the authors' own prior co-design work, which is directly relevant.\n\nWho this is for: people working on co-design and robust control of legged robots will find a useful tool description and a clear cautionary example. The paper deserves serious referee time — the idea is promising and the gaps are addressable — but as written the central claim is not established. I would send it to review, asking for a reframing around the largest exponent or explicit disturbance-rejection benchmarks, plus released artifacts.","headline":"A genuinely new differentiable Lyapunov-exponent objective for robot co-design, but the paper's scalar metric Lλ does not actually guarantee the non-divergence it is claimed to measure.","tokens_in":9950,"tokens_out":3214,"would_cite":false,"duration_ms":36404,"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":"The signed sum of the Lyapunov exponents of a closed-loop robot is a robustness metric that differentiable simulation can minimize.","keywords":["Lyapunov exponents","robustness metric","differentiable simulation","co-design","contact-rich dynamics","limit cycles","legged locomotion","gradient-based optimization"],"falsifier":"Run a sweep of gains and designs, rank them by $L_\\lambda$, and physically perturb the best and worst: if any controller with more negative $L_\\lambda$ rejects a finite push, terrain drop, or motor-torque noise worse than a less negative one, the proxy is falsified as a general robustness metric. A concrete version is to apply an impulsive side force to the optimized manipulator of Table I and measure settling time and success rate against a nominal design; the ranking by $L_\\lambda$ predicts the optimized design wins every such test.","tokens_in":1764,"feed_emoji":"🤖","tokens_out":1628,"duration_ms":104240,"temperature":0.7,"pith_summary":"Robustness to perturbations is hard to define, let alone optimize. The paper argues that a single number, the signed sum $L_\\lambda = \\mathbf{1}^\\top \\lambda$ of the Lyapunov exponents of the closed-loop system, captures how strongly nearby trajectories converge, and that making this number more negative improves the robot's ability to reject disturbances. The method computes $\\lambda$ by differentiating a long simulation rollout: at each step, the Jacobian $\\mathrm{d}\\Phi(x_i)$ of the state transition is obtained from a differentiable physics simulator, the spectrum is extracted from an SVD of $\\mathrm{d}\\Phi^\\top\\mathrm{d}\\Phi$, and the time average of the log singular values gives $\\lambda$. Because the whole pipeline is differentiable, $L_\\lambda$ can be minimized together with task costs by gradient-based co-optimization of hardware and policy parameters. This matters because, if valid, it turns the vague goal of robustness into a concrete objective usable for analysis, design, and control tuning in high-degree-of-freedom, contact-rich robots.","feed_headline":"Summing Lyapunov exponents yields a trainable robot robustness score","feed_subtitle":"Differentiable simulation turns the contraction of nearby trajectories into a co-design objective for legs and arms.","key_machinery":"The load-bearing object is the finite-horizon Lyapunov spectrum computed from a differentiable rollout. Let $\\Phi(x_i)$ be one simulation step and $\\mathrm{d}\\Phi(x_i)$ its Jacobian; the paper estimates the exponents as\n$$\n\\$\\lambda$ = \\frac{1}{2N\\$\\Delta$ t}\\sum_{i=1}^N \\log(\\Sigma_i),\n$$\nwhere $\\Sigma_i$ are the singular values of $\\mathrm{d}\\Phi(x_i)^\\top\\mathrm{d}\\Phi(x_i)$ — an SVD variant chosen so that floating-base robots with non-square Jacobians remain tractable. The robustness score is the signed sum $L_\\lambda = \\mathbf{1}^\\top\\lambda$, which the paper identifies with the degree of hypervolume contraction in phase space: negative $L_\\lambda$ means volumes shrink and trajectories converge. Differentiability of the rollout and of the SVD is what lets this scalar be optimized by first-order methods for hardware parameters and policy gains alike.","core_discovery":"On the paper's own terms, the central claim is that the signed sum of the Lyapunov exponents, $L_\\lambda = \\mathbf{1}^\\top \\lambda$, is an effective robustness metric for robotic systems, naturally covering fixed points and limit cycles, and that minimizing it via differentiable simulation improves perturbation rejection. Evidence is reported across three case studies: co-optimizing a two-link planar manipulator's link lengths, concentrated masses, and PD gains moves $L_\\lambda$ from $-2.5\\cdot 10^{-3}$ to $-1.3\\cdot 10^{-2}$ and makes 100 randomly started trajectories converge more steadily; for a 16-joint spider falling from 1 m, the metric ranks gain combinations, with heavier damping and lower stiffness appearing more robust; and for a crawling quadruped with 8 actuated joints and a floating base, optimizing a 9-parameter sinusoidal policy yields $L_\\lambda = -0.61$ versus $-0.56$ for the baseline. The paper presents this as the first gradient-based use of Lyapunov-exponent theory on a contact-rich, high-dimensional legged system.","pith_inferences":["Editorial inference: because $L_\\lambda$ rewards phase-space volume contraction only, unweighted minimization may favor overdamped, energy-dissipating designs; a testable extension is to include a task-performance weight or a finite-disturbance validation loop in the loss.","Editorial inference: the asymptotic average may not reflect finite-time robustness, so finite-time Lyapunov exponents or explicit disturbance injection would be a natural complement to this metric.","Editorial inference: for non-smooth contact dynamics the ergodic assumptions behind classical Lyapunov exponents may be questionable, making the quantity best read as a practical finite-horizon indicator rather than a rigorous global characteristic."],"forward_implications":["Co-design of robot morphology and controller can be posed as minimizing one scalar, $L_\\lambda$, alongside task costs, with gradients obtained from a differentiable simulator.","The metric extends naturally to limit cycles, so periodic locomotion policies can be robustified without first stabilizing an equilibrium.","The same $L_\\lambda$ value can be used to rank controllers or designs offline, as in the spider study where gain maps reveal a damping-versus-stiffness trade-off.","Because the full spectrum is computed rather than a single exponent, the method can in principle distinguish chaotic divergence in some directions from contraction in others.","Long-horizon rollouts used to estimate $\\lambda$ may degrade gradient quality in contact-rich loco-manipulation, a limitation the paper explicitly acknowledges."],"supporting_citations":[{"why":"Supplies the multiplicative ergodic theorem, the theoretical foundation for Lyapunov exponents as limits of Jacobian products.","marker":"[25]"},{"why":"Provides the standard numerical scheme for spectra that the paper adapts, via SVD, to non-square Jacobians.","marker":"[24]"},{"why":"Establishes differentiability of the SVD, which keeps the Lyapunov computation inside the gradient path.","marker":"[27]"},{"why":"Provides the differentiable physics engine used for forward rollouts, contact interactions, and Jacobian extraction.","marker":"[30]"},{"why":"Provides automatic differentiation and parallel map operations used to compute $\\mathrm{d}\\Phi$ and backpropagate through the rollout.","marker":"[31]"},{"why":"Earlier speculation that Lyapunov exponents could guide robot hardware optimization, motivating the application.","marker":"[16]"},{"why":"Prior use of Lyapunov exponents for biped disturbed standing, connecting the metric to legged robustness.","marker":"[17]"},{"why":"Liouville's theorem grounds the claim that volume-conserving Hamiltonian systems have $L_\\lambda = 0$, giving the sign interpretation.","marker":"[29]"}],"fun_headline_variants":["Lyapunov exponents become a trainable robot robustness metric","Minimize Lyapunov sum to make robots more robust","Differentiable Lyapunov metric optimizes robot designs","First gradient-based Lyapunov optimization for legged robots","Robustness by Lyapunov sum: a new co-design objective"],"cache_read_input_tokens":11904,"weakest_assumption_plain":"The approach treats the more-negative-is-more-robust relation as axiomatic: no proof, benchmark, or quantitative link to finite disturbances, actuator noise, or task performance is given, so optimizing $L_\\lambda$ could simply produce overdamped or degenerate designs that shrink phase-space volume without rejecting the disturbances that matter.","fun_headline_variants_meta":{"raw":{"variants":["Lyapunov exponents become a trainable robot robustness metric","Minimize Lyapunov sum to make robots more robust","Differentiable Lyapunov metric optimizes robot designs","First gradient-based Lyapunov optimization for legged robots","Robustness by Lyapunov sum: a new co-design objective"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1532,"prompt_tokens":884,"completion_tokens":648,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":565}},"tokens_in":500,"tokens_out":648,"duration_ms":6568,"temperature":1.0,"reasoning_tokens":565,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:23:44.244716+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a sweep of gains and designs, rank them by $L_\\lambda$, and physically perturb the best and worst: if any controller with more negative $L_\\lambda$ rejects a finite push, terrain drop, or motor-torque noise worse than a less negative one, the proxy is falsified as a general robustness metric. A concrete version is to apply an impulsive side force to the optimized manipulator of Table I and measure settling time and success rate against a nominal design; the ranking by $L_\\lambda$ predicts the optimized design wins every such test.","supporting_citations":[{"cited_title":"A multiplicative ergodic theorem. Lyapunov char- acteristic numbers for dynamical systems","cited_arxiv_id":null,"evidence_quote":"Supplies the multiplicative ergodic theorem, the theoretical foundation for Lyapunov exponents as limits of Jacobian products."},{"cited_title":"Nonlinear Dynamics: A Concise Introduction Interlaced with Code","cited_arxiv_id":null,"evidence_quote":"Provides the standard numerical scheme for spectra that the paper adapts, via SVD, to non-square Jacobians."},{"cited_title":"Mujoco: A physics engine for model-based control","cited_arxiv_id":null,"evidence_quote":"Provides the differentiable physics engine used for forward rollouts, contact interactions, and Jacobian extraction."},{"cited_title":"JAX: composable transformations of Python+NumPy programs","cited_arxiv_id":null,"evidence_quote":"Provides automatic differentiation and parallel map operations used to compute $\\mathrm{d}\\Phi$ and backpropagate through the rollout."},{"cited_title":"Applications of Chaotic Dynamics in Robotics","cited_arxiv_id":null,"evidence_quote":"Earlier speculation that Lyapunov exponents could guide robot hardware optimization, motivating the application."},{"cited_title":"On stabilization of bipedal robots during dis- turbed standing using the concept of Lyapunov exponents","cited_arxiv_id":null,"evidence_quote":"Prior use of Lyapunov exponents for biped disturbed standing, connecting the metric to legged robustness."},{"cited_title":"Liouville’s theorem","cited_arxiv_id":null,"evidence_quote":"Liouville's theorem grounds the claim that volume-conserving Hamiltonian systems have $L_\\lambda = 0$, giving the sign interpretation."}],"review_version":1}