{"id":"9cda29cf-7e57-40e7-a34e-2534649fceab","arxiv_id":"2505.19521","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Measurement-aware control barrier functions are re-framed through a fiber bundle geometry for Neural ODE dynamics learning, but the claimed convergence and safety guarantees rely on an oracle update and a faulty union bound.","lead":"This paper proposes a geometric framework in which sensor measurements form a fiber bundle over a robot's state space, used to build measurement-aware barrier functions and learn dynamics with Neural ODEs. If the framework worked, robots could learn safely from local, noisy sensors without global environment knowledge, which matters for real-world autonomy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1's convergence proof is for an oracle update (Eq. 12) containing the unknown true dynamics f, while the implemented algorithm (Eq. 14) never sees f; the exponential bound is not established for the proposed method.","rationale":"The reader's weakest assumption is that Theorem 4.1's convergence proof uses an update containing the unknown f, while Eq. (14) minimizes a data loss; I agree this is the most load-bearing flaw. The central claim is specifically that the learned estimate converges with the stated exponential bound. A theorem proved for dot(hat f) = -L1(hat f - f) cannot be a guarantee about a network trained by Eq. (14) unless an equivalence map is provided; none is given. I also noted the safety proof's union-bound error and the unjustified positive lower bound b_min in Appendix A, but I do not make those the primary objection because the oracle mismatch alone severs the main learning guarantee from the implemented algorithm. The experiments are extensive and the code link is useful, but they measure task success and constraint satisfaction rather than ||hat f - f||_E, so they do not substitute for the missing proof. Given the structural gap, the existing REJECT verdict is appropriate; no adjustment is needed.","tokens_in":20853,"tokens_out":9439,"duration_ms":82200,"concrete_test":"Attempt to re-derive the first inequality of Theorem 4.1 starting from Eq. (14) instead of Eq. (12): set hat f = hat f_theta, use dot theta = -Lambda nabla_theta T(hat f_theta,D), and try to upper-bound d/dt ||hat f_theta - f||_E^2 by -lambda ||hat f_theta - f||_E^2 + delta_v^2 using only D and the neural network structure. Record which step requires an oracle term containing f, such as an inner product with f - hat f_theta that is not recoverable from noisy targets. A companion numerical check: train Eq. (14) on one Section 5 task and log ||hat f_theta - f|| over time against the claimed c1 exp(-lambda1 t) + c2 delta_v; if the measured error does not decay at the predicted rate or plateaus above the delta_v level, the oracle mismatch is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central learning claim is the first inequality of Theorem 4.1: ||hat f - f||_E <= c1 exp(-lambda1 t) + c2 delta_v. Appendix B proves it by differentiating V = (1/2)||hat f - f||_E^2 along dot(hat f) = -L1(hat f - f) (Eq. 12). That update law contains the unknown true dynamics f as a target signal; it is an oracle equation. The implemented rule in Eq. (14) updates parameters theta by gradient descent on a data loss using measured state derivatives dot x_i; no term in it is f itself. No mapping from theta to hat f is given under which Eq. (12) holds, nor is a bound on dot(hat f_theta) + L1(hat f_theta - f) supplied. The Lyapunov calculation therefore proves convergence of a different, oracle process. Even for that oracle process, Eq. (38) adds a ||hat f - f|| delta_v term with no source: if dot(hat f) = -L1(hat f - f), then dot V = -<hat f - f, L1(hat f - f)>, with no measurement noise. The c2 delta_v residual is inserted rather than derived. This is structural: the theorem's premise concerns what information the learner receives, not a parameter setting. The safety proof is separately invalid (Appendix A's union-bound series evaluates to exp(c1 Delta t - c2/delta_v^2), which is at least 1 under the stated condition c1 Delta t >= c2/delta_v^2), but the oracle gap alone breaks Theorem 4.1's main convergence statement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a geometric framework in which measurement uncertainty induces a fiber bundle over the state space, used to define measurement-adapted control barrier functions (mCBFs) and a Neural ODE-based dynamics learner. The main theoretical claims are Theorem 3.1, a probabilistic safety guarantee of the form P(x(t) in S0 for all t >= 0) >= 1 - exp(-c/delta_v^2), and Theorem 4.1, an exponential convergence bound for the learned dynamics with a measurement-noise residual together with a safety guarantee during learning. The paper also presents simulation results on soft-body, manipulator, and quadrotor tasks, plus ablations and cross-domain experiments.","tokens_in":21225,"tokens_out":5628,"duration_ms":48319,"significance":"If established, the framework would be a noteworthy unification of geometric control and learning: the safety guarantee degrades gracefully with measurement noise, and the convergence bound is parameterized by sensing quality. The paper ships a public implementation and reports extensive experiments and ablations, which is a strength. However, the theoretical results are the main advertised contribution, and the proofs contain load-bearing gaps: the safety proof's final union bound evaluates to a trivial bound, and the convergence proof analyzes an oracle update that uses the unknown true dynamics rather than the implemented parameter update. Until these issues are resolved, the theoretical significance claimed in the paper is not supported.","major_comments":[{"comment":"The union bound is evaluated incorrectly. The series sum_{k=0}^infty exp(-c2/delta_v^2) (1 - exp(-c1 Delta t))^k equals exp(c1 Delta t - c2/delta_v^2), which is at least 1 under the stated condition c1 Delta t >= c2/delta_v^2. Therefore Eq. (35) does not imply the claimed probability bound, and Theorem 3.1 is unproven. In addition, Lemma A.4 is asserted with no derivation; it requires a temporal decorrelation property of the events A_t, but these events depend on the state trajectory, so independence or exponential mixing is not automatic from sub-Gaussian measurement noise.","section":"Appendix A, Eq. (35)"},{"comment":"Theorem 4.1's convergence analysis is performed for the update dot(hat f) = -L1(hat f - f), whose right-hand side contains the unknown true dynamics f. The implemented rule in Eq. (14) updates parameters theta by gradient descent on a data loss and is never shown to realize Eq. (12); no relation between theta and hat f is given, and no bound on dot(hat f_theta) + L1(hat f_theta - f) is supplied. The exponential convergence bound therefore applies to an oracle process, not to the proposed algorithm.","section":"Section 4.1-4.3, Eqs. (12) and (14)"},{"comment":"The Lyapunov derivative contains the term ||hat f - f||_E delta_v, but from dot(hat f) = -L1(hat f - f) one obtains dot(V) = -<hat f - f, L1(hat f - f)> exactly. The measurement-noise residual c2 delta_v is inserted rather than derived, so even the oracle convergence bound with a noise-dependent residual is not established.","section":"Appendix B, Eq. (38)"},{"comment":"The proof of the second inequality in Theorem 4.1 assumes b(x(t), h(x(t))) >= b_min for all t along the learning rollout. Lemma A.1 establishes a similar statement only under perfect measurements (v = 0) and without the policy/learning coupling; the effect of the learned dynamics and the actual measurement noise on the mCBF value is not analyzed. Consequently, the safety guarantee during learning, P(x(t) in S0) >= 1 - exp(-c3/delta_v^2), is not proven, and the statement's quantifier over t is ambiguous.","section":"Appendix B, Eq. (44)"}],"minor_comments":[{"comment":"The connection formula uses pi_*^{-1} and the term K(x)(y - h(x)); these are not defined as vector fields on E, so the expression is not mathematically well-formed as written.","section":"Section 3.2, Eq. (3)"},{"comment":"The mCBF condition omits explicit dependence on the control input u in L_f b; the infimum over u is taken, but f(x,u) and the chosen control law should be specified.","section":"Section 3.6, Eq. (9)"},{"comment":"The sensor noise model ||v_i|| <= alpha ||x - x_i|| is state-dependent, while the theory assumes a global bound ||v|| <= delta_v; the relation between the two should be stated.","section":"Section 5.2"},{"comment":"The additional environments are described as motivated by real datasets but are themselves simulations; the text should clarify that no real sensor data are used.","section":"Appendix D"},{"comment":"The theorems state constants c, c1, c2, c3, lambda1 without specifying their dependence on system parameters; the proofs should state the relevant dependencies for the claims to be checkable.","section":"Theorems 3.1 and 4.1"}],"recommendation":"reject","confidential_remarks":"I concur with the stress-test analysis: the oracle-update issue and the union-bound error are independent and both undermine the central theoretical claims. The experimental results cannot compensate because the paper's stated contribution is the theoretical guarantee. I would not rule out a future resubmission if the proofs are repaired and the learning update is actually analyzed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a serious empirical paper wearing a theory coat that doesn't fit. The three simulation tasks are sensible, the ablations are thorough, and the reported gains over baselines are plausible engineering results. But the theoretical guarantees, which are the main selling point, fail on close reading.\n\nTheorem 3.1's proof in Appendix A ends with a geometric series that evaluates to exp(c1 Δt − c2/δ_v^2), which is ≥1 under the paper's own condition c1 Δt ≥ c2/δ_v^2. The bound doesn't follow. Lemma A.4 also just assumes temporal decorrelation without an argument. Theorem 4.1 is worse: the convergence proof differentiates a Lyapunov function along Eq. (12), dot(hat f) = −L1(hat f − f). That update law has the true dynamics f on the right-hand side. The implemented rule, Eq. (14), is gradient descent on a data loss and never sees f. So the exponential convergence is proven for an oracle process, not for the proposed algorithm. The extra δ_v term in Eq. (38) also appears from nowhere. These are load-bearing, not cosmetic.\n\nWhat's new? The mCBF conditions reduce essentially to a robust CBF with a Lipschitz measurement margin; Appendix G spells out b(x,y)=b0(x,h(x))−L_b||y−h(x)||, which is close to Taylor & Ames's adaptive safety. The fiber-bundle language doesn't add mathematical content that I can see. The experiments do show consistent improvements, and the code link and hyperparameter detail help reproducibility, though there is no commit hash. The noise model is also internally inconsistent: y=h(x)+v with |v|≤δ_v, but the proof uses a sub-Gaussian tail bound.\n\nBottom line: the empirical part is worth a look for people building measurement-aware safety filters, but the theoretical claims should not be relied on. I'd send it to reviewers only to confirm the proof failures; as a contribution it needs major restructuring. If the authors decouple the empirical recipe from the false guarantees and either fix or drop the theorems, there might be a usable paper.","headline":"Well-executed empirical paper whose two main theorems don't prove what the claims say — the convergence proof runs on an oracle update and the safety proof's final bound doesn't follow — so it's not citable as theory.","tokens_in":21803,"tokens_out":3696,"would_cite":false,"duration_ms":32265,"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":"This paper claims that a robot's locally noisy sensor readings are not noise to be filtered out but a geometric structure — a fiber bundle over state space — that turns each measurement into an adaptive safety certificate and drives…","keywords":["fiber bundle","measurement-adapted control barrier functions","safe learning control","dynamics learning","Neural ODE","measurement uncertainty","safety certificates","local sensing"],"falsifier":"Train the implemented rule of Eq. (14) on any of the three tasks while measuring the bundle-norm error $\\|\\hat f - f\\|_E$ against training time and noise bound $\\delta_v$: the paper predicts decay like $c_1 e^{-\\lambda_1 t} + c_2 \\delta_v$, so a plateau well above $c_2 \\delta_v$ or non-exponential decay would contradict that prediction for the implemented algorithm. Independently, count constraint violations over many seeds at several noise levels: the safety claim predicts a rate scaling like $\\exp(-c/\\delta_v^2)$, so a log-linear plot against $1/\\delta_v^2$ that is markedly flatter than the predicted slope would contradict Theorem 3.1. The paper itself notes the implementation is limited for highly stochastic dynamics, the regime where the sub-Gaussian-noise assumption behind the exponential bound fails; a test with heavy-tailed noise would probe that boundary.","tokens_in":20546,"feed_emoji":"🤖","tokens_out":20191,"duration_ms":165371,"temperature":0.7,"pith_summary":"This paper aims to establish that a robot can learn unknown dynamics and guarantee safety using only local, noisy measurements, by treating measurement uncertainty as geometry rather than as an external disturbance. The central construction is a fiber bundle over the state space whose fiber above each state contains every measurement the sensors could return for that state; on this bundle the paper defines measurement-adapted control barrier functions, safety certificates whose margins tighten automatically where sensing is poor. With such a certificate, the paper proves that the probability of staying inside the safe set for all future time is at least $1 - \\exp(-c/\\delta_v^2)$, and that a Neural ODE trained within this geometry converges to the true dynamics exponentially fast, up to a residual proportional to the measurement noise bound $\\delta_v$. If these claims are right, robots could learn and act safely with cheap, imperfect sensors and no global description of the constraints, and both safety and learning quality would degrade in a predictable way as sensing worsens.","feed_headline":"Local, noisy sensors can certify robot safety while it learns","feed_subtitle":"Bundle geometry turns fuzzy local readings into safety certificates whose margins tighten automatically as sensing sharpens.","key_machinery":"The load-bearing object is the measurement-induced fiber bundle $\\pi: E = M \\times Y \\to M$ with fiber $\\pi^{-1}(x) = \\{(x, y) : y = h(x) + v,\\ \\|v\\| \\le \\delta_v\\}$, together with a connection $\\nabla$ that couples state evolution with measurement evolution through a measurement-feedback gain operator $K$. On this bundle the paper defines the measurement-adapted control barrier function (mCBF) $b: E \\to \\mathbb{R}$, whose three defining conditions do the real work: the implication $b(x,y) \\ge 0 \\Rightarrow x \\in S_0$ ties the certificate to the safe set; the Lie-derivative condition $\\inf_u[L_f b + (L_g b)w + \\alpha(b)] \\ge 0$ provides forward invariance along the controlled flow; and the Lipschitz condition $|b(x,y_1) - b(x,y_2)| \\le L_b d_Y(y_1, y_2)$ converts bounded measurement error into bounded erosion of the safety margin. The connection propagates that margin along trajectories, and the sub-Gaussian tail of the measurement noise turns the erosion into the probability bound $1 - \\exp(-c/\\delta_v^2)$. For learning, the mechanism is the bundle-valued update on the estimate $\\hat f$ combined with the uncertainty-weighted data loss of Eq. (14), whose inverse-covariance weights $\\Sigma_i^{-1}$ feed the measurement geometry directly into the gradient.","core_discovery":"The paper's claim is that the pairing of a state manifold with its uncertain measurements is itself a fiber bundle $\\pi: E = M \\times Y \\to M$ with fiber $\\pi^{-1}(x) = \\{(x,y) : y = h(x) + v,\\ \\|v\\| \\le \\delta_v\\}$, and that this induced geometry carries the entire argument. A measurement-adapted control barrier function $b: E \\to \\mathbb{R}$ is a bundle map that certifies safety through $b(x,y) \\ge 0 \\Rightarrow x \\in S_0$, satisfies the forward-invariance condition $\\inf_{u \\in U}[L_f b + (L_g b)w + \\alpha(b)] \\ge 0$ along admissible controls, and is $L_b$-Lipschitz in the measurement direction. Theorem 3.1 states that whenever such a certificate starts nonnegative, every admissible noise sequence satisfies $P(x(t) \\in S_0 \\text{ for all } t \\ge 0) \\ge 1 - \\exp(-c/\\delta_v^2)$. Theorem 4.1 states that an estimator trained through the bundle's uncertainty-weighted loss satisfies $\\|\\hat f - f\\|_E \\le c_1 e^{-\\lambda_1 t} + c_2 \\delta_v$: the learned model converges exponentially to the true dynamics with a noise-limited floor, while the safety bound continues to hold with a constant $c_3$. Together these assert that sensing quality sets the achievable rate of both safety and learning.","pith_inferences":["One consequence the authors leave implicit is a sensor-economics trade: any change to the measurement map or to sensor placement that shrinks the product $L_b \\delta_v$ directly tightens the safety bound and lowers the learning floor, so choosing sensors becomes an optimization on the bundle geometry rather than a filtering problem.","The safety theorem is stated for certificates built on the true dynamics $f$; extending it to a controller that runs on the learned $\\hat f$ requires showing the mCBF conditions survive the estimation error, for instance by feeding Theorem 4.1's bound back into the invariance condition, a loop the paper does not close.","The bound's dependence on sensing quality is sharp enough to test: plotting constraint-violation frequency against $1/\\delta_v^2$ on a log scale in the three simulated tasks should show the predicted $\\exp(-c/\\delta_v^2)$ decay, a stronger fingerprint than a generic 'more noise, more violations' trend."],"forward_implications":["Safety becomes location-adaptive: where sensors are accurate the certificate relaxes, where they are uncertain it tightens, and a single certificate works without a global map of the obstacle field.","The probability of ever leaving the safe set is bounded by $\\exp(-c/\\delta_v^2)$, so the worst-case violation rate decays faster than any polynomial as sensing sharpens.","The learned model approaches the true dynamics in the bundle norm at an exponential rate, with a floor $c_2 \\delta_v$ set by measurement noise; learning quality is therefore limited by sensing quality rather than by the neural architecture.","Because no global constraint knowledge is assumed, the same certificate transfers to new obstacle configurations, which the paper supports with a reported 96.3% success rate across the worm, arm, and quadrotor tasks.","Policy updates can be projected onto the certificate's safe-policy set, which the paper offers as a route for reinforcement learning with partial, uncertain observations."],"supporting_citations":[{"why":"Supplies the control-barrier-function formulation with quadratic programs whose forward-invariance condition the mCBF definition extends to uncertain measurements.","marker":"Ames et al., 2016"},{"why":"Foundational CBF framework providing formal constraint-satisfaction guarantees for known dynamics, the guarantee the paper generalizes to the measurement-adapted setting.","marker":"Ames et al., 2019"},{"why":"Introduces Neural ODEs, the continuous-time architecture used to represent and train the learned dynamics estimate.","marker":"Chen et al., 2018"},{"why":"Origin of the infinitesimal connection on fiber bundles that the paper places over the state space to couple dynamics with measurement evolution.","marker":"Ehresmann, 1950"},{"why":"Canonical reference for the differential-geometric connections and covariant derivatives used to define the bundle structure and its safety maps.","marker":"Kobayashi & Nomizu, 1996"},{"why":"Earlier approach to learning control under measurement noise that the paper positions its geometric treatment against.","marker":"Berkenkamp et al., 2017"},{"why":"Provides the simulation engine used for all three reported experimental tasks.","marker":"Xian et al., 2023"}],"fun_headline_variants":["Bundle geometry turns fuzzy sensing into tight safety certificates","Measurement-aware barriers guarantee safe learning despite noise","Safety margins tighten as sensing sharpens: bundle geometry","Fuzzy measurements become safety guarantees via bundles","Local uncertain readings still yield safe learning guarantees"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learning-convergence guarantee of Theorem 4.1 is proved for the update law in Eq. (12), whose right-hand side contains the true, unknown dynamics $f$, while the training rule actually run in the experiments, Eq. (14), learns from data that never reveal $f$; unless the two rules are shown to behave identically, the exponential convergence bound belongs to the proof's idealized process rather than to the implemented algorithm.","fun_headline_variants_meta":{"raw":{"variants":["Bundle geometry turns fuzzy sensing into tight safety certificates","Measurement-aware barriers guarantee safe learning despite noise","Safety margins tighten as sensing sharpens: bundle geometry","Fuzzy measurements become safety guarantees via bundles","Local uncertain readings still yield safe learning guarantees"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001274,"raw_usage":{"total_tokens":5249,"prompt_tokens":1020,"completion_tokens":4229,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":4160}},"tokens_in":636,"tokens_out":4229,"duration_ms":29528,"temperature":1.0,"reasoning_tokens":4160,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:13:02.673922+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the implemented rule of Eq. (14) on any of the three tasks while measuring the bundle-norm error $\\|\\hat f - f\\|_E$ against training time and noise bound $\\delta_v$: the paper predicts decay like $c_1 e^{-\\lambda_1 t} + c_2 \\delta_v$, so a plateau well above $c_2 \\delta_v$ or non-exponential decay would contradict that prediction for the implemented algorithm. Independently, count constraint violations over many seeds at several noise levels: the safety claim predicts a rate scaling like $\\exp(-c/\\delta_v^2)$, so a log-linear plot against $1/\\delta_v^2$ that is markedly flatter than the predicted slope would contradict Theorem 3.1. The paper itself notes the implementation is limited for highly stochastic dynamics, the regime where the sub-Gaussian-noise assumption behind the exponential bound fails; a test with heavy-tailed noise would probe that boundary.","supporting_citations":[{"cited_title":"T., Rubanova, Y., Bettencourt, J., and Duvenaud, D","cited_arxiv_id":null,"evidence_quote":"Introduces Neural ODEs, the continuous-time architecture used to represent and train the learned dynamics estimate."},{"cited_title":"Les connexions infinit \\'e simales dans un espace fibr \\'e diff \\'e rentiable","cited_arxiv_id":null,"evidence_quote":"Origin of the infinitesimal connection on fiber bundles that the paper places over the state space to couple dynamics with measurement evolution."},{"cited_title":"and Nomizu, K","cited_arxiv_id":null,"evidence_quote":"Canonical reference for the differential-geometric connections and covariant derivatives used to define the bundle structure and its safety maps."},{"cited_title":"Safe model-based reinforcement learning with stability guarantees","cited_arxiv_id":null,"evidence_quote":"Earlier approach to learning control under measurement noise that the paper positions its geometric treatment against."},{"cited_title":"Towards Generalist Robots: A Promising Paradigm via Generative Simulation","cited_arxiv_id":"2305.10455","evidence_quote":"Provides the simulation engine used for all three reported experimental tasks."}],"review_version":1}