{"id":"0d36c7cf-4183-4851-969b-024fc0dff72f","arxiv_id":"2504.19989","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HJRNO learns a mapping from the signed-distance function of an obstacle to the infinite-horizon reachability value function, achieving millisecond inference on held-out obstacles and reduced-state system settings.","lead":"A team at Simon Fraser University trained neural operators (FNO and TNO) to predict backward reachable tubes, the unsafe states an autonomous system must avoid, from obstacle geometry in milliseconds. The method generalizes to new obstacle shapes in 2D test cases, but its treatment of state variables as constants limits the claim to full dynamic systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reduction of velocity and heading to constant hyperparameters is unvalidated; for the dynamic Dubins car these are evolving states, so reported errors may be for a surrogate 2D problem rather than the true BRT.","rationale":"I read the paper as a supervised operator-learning benchmark on solver-generated data. The learned mapping and fast inference are plausible for the reduced 2D problems, and comparing FNO and TNO is a useful contribution. The load-bearing gap is the reduction of state dimensions to hyperparameters. This is exactly the reader's weakest assumption. The manuscript itself flags velocity dependence in Section 5.2 while simultaneously assuming approximate constancy in Section 5, so the concern lands. The reported low errors do not establish accuracy for the true full-order BRTs because no full-order ground truth is provided. I would keep the verdict CONDITIONAL: the empirical claims need full-order validation before the generalization claim can be accepted. I see no reason to move to REJECT because the reduced-problem results are internally coherent and the missing comparison is a well-defined additional experiment rather than a demonstrated failure.","tokens_in":9906,"tokens_out":6368,"duration_ms":66370,"concrete_test":"Use the Stanford hj_reachability JAX solver to compute the full-order 4D BRT for the dynamic Dubins car on 5-10 test obstacle scenarios from the paper (grid 50^4, converged horizon). Slice the resulting V∞(x1,x2,v,θ) at the exact (v0,θ0) encoded as hyperparameters in HJRNO, and compare the slice to HJRNO's 2D output: compute relative L2 error and BRT-boundary Hausdorff distance. Separately, measure the full-order value function's variation in v and θ across the slice; if the range exceeds roughly 10% of the value magnitude, the 'approximately constant' assumption in Section 5 fails. Repeat for the Velocity-Dependent setup, where V depends on v by construction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the learned operator maps the initial value function of the true system to its converged BRT value function. Section 5 (paragraph 2) reduces the state by declaring that V∞ 'typically varies only with the spatial coordinates... while remaining approximately constant along other dimensions such as velocity or heading,' and Eq. (26) encodes v and θ as constant hyperparameters h. For the dynamic Dubins car, however, v and θ are states with dynamics v˙=u1 and θ˙=v u2 (Eq. 7). A 2D slice at fixed h is the solution of a different, lower-dimensional system: the value function for a frozen velocity and heading, not the value function of the full-order system, because those hyperparameters are not held constant in time. The Velocity-Dependent experiment makes this concrete: it intentionally makes the obstacle radius grow with v, so V∞ varies with v by construction; encoding v as a constant input cannot produce a full 4D value function. No comparison against full-order ground truth is given. The only full-state result, Air3D, shows degradation (FNO 0.028→0.050, TNO 0.037→0.15), which points to the full problem being qualitatively harder rather than validating the reduced model. Thus the reported 0.002-0.037 errors may only be errors on a 2D surrogate, and the abstract's claim of generalization across system dynamics is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HJRNO, a neural-operator approach for computing backward reachable tubes (BRTs) in Hamilton-Jacobi reachability analysis. Using Fourier Neural Operators (FNO) and Transformer-based Neural Operators (TNO), the authors train models to map an initial value function (equivalently, the obstacle geometry, optionally augmented by scalar parameters) to the infinite-horizon converged value function. The training data are generated offline by a standard HJI PDE solver using the hj_reachability library. The paper reports relative L2 test errors between 0.002 and 0.037 on several obstacle-shape scenarios, including Air3D collision avoidance, single and multiple obstacles, indoor environments, velocity-dependent obstacles, and parametric control limits, with inference times in the millisecond range. It also demonstrates zero-shot super-resolution, where models trained at one grid resolution generalize to a coarser or finer resolution without retraining.","tokens_in":10189,"tokens_out":5508,"duration_ms":58337,"significance":"If the claims hold, the operator-learning viewpoint is a natural and potentially impactful way to amortize HJR computations, because the input is a function (the initial value field) rather than a scalar parameter set. The use of held-out obstacle shapes guards against simple memorization, and the low reported errors on shape generalization, together with the speed advantage, would be useful for real-time robotics safety filtering. The zero-shot super-resolution comparison with CNNs is a positive contribution that concretely illustrates the discretization-invariance benefit of neural operators. However, the evidence for generalization across 'system dynamics' is currently much weaker than the evidence for shape generalization, because non-spatial state dimensions are removed from the problem in the experiments. The paper's central contribution is therefore best assessed as an operator approximation for two-dimensional value functions, with the claims about full-order system dynamics needing substantial additional support.","major_comments":[{"comment":"The paper's main claim of generalization across system dynamics rests on encoding state variables as constant hyperparameters. For the dynamic Dubins car (Eqs. 7a-7d), velocity v and heading θ are evolving states, not problem hyperparameters, and the value function of the full 4D system is not generally recoverable from 2D slices at fixed v and θ. The Velocity-Dependent experiment makes this concrete: because the obstacle radius is intentionally made a function of v, the converged value function V∞ varies with v by construction, so encoding v as a constant input function h cannot represent the full value function. As currently presented, the reported errors in Tables 1 and 2 appear to describe a two-dimensional surrogate problem rather than the true BRT of the stated dynamics. The authors should either provide full-order ground-truth comparisons for the velocity-dependent and parametric-input experiments, or explicitly restrict the claims of the paper to the reduced model.","section":"Section 5, paragraph 2; Eq. (26); Section 2.1 (Dynamic Dubins car)"},{"comment":"The only full-state result reported in the paper is Air3D, and it shows substantial degradation when the heading state is included (FNO test error 0.028 to 0.050; TNO test error 0.037 to 0.15). This is consistent with the concern that the low errors in Tables 1 and 2 are achieved on a reduced two-dimensional problem rather than the actual system. The degradation also indicates that the full-order problem is qualitatively harder, which further undermines the abstract's claim that the method 'generalizes effectively across varying system dynamics.' The authors should include full-order experiments in the main evaluation, or revise the abstract and introduction to describe the contribution as applying to value functions on a reduced state space.","section":"Section 5.1 (Air3D) and the full-state Air3D paragraph"},{"comment":"No error bars or standard deviations are reported for the relative L2 errors, and no baselines are given against prior learning-based reachability methods such as DeepReach or parameter-conditioned HJR solvers. Since training is stochastic and the datasets are small (50-300 samples), a single run does not establish that the reported errors are robust. At minimum, the authors should report the mean and standard deviation over several training seeds, and compare against a standard supervised-learning baseline on the same input-output pairs, so that the claimed advantages of the neural-operator formulation can be assessed.","section":"Tables 1 and 2"}],"minor_comments":[{"comment":"The text says 'dynamics Dubins car'; this should read 'dynamic Dubins car' for consistency with the terminology introduced in Section 2.","section":"Section 5.2"},{"comment":"The figure labels contain the typo 'Initual Value'; this should be 'Initial Value'.","section":"Figures 8, 9, 11 and Section 5"},{"comment":"The example 'the maximum acceleration limit in the dynamic Dubins car' is presented as a hyperparameter, but the preceding discussion of reducing state variables to hyperparameters could be misread; please clarify that control-limit bounds are hyperparameters, while v and θ are not.","section":"Section 4.1"},{"comment":"The 'Data Size' of 2.4 GB for a 100-sample, 50x50 two-dimensional dataset is inconsistent with the paper's own estimate that a full 50^4 array is about 25 MB; please clarify what is being stored.","section":"Tables 1 and 2"},{"comment":"The attention formula mixes spatial and feature indices in a way that is hard to follow; a standard multi-head attention definition would be clearer for the reader.","section":"Section 3.2.2, Eq. (21)"}],"recommendation":"major_revision","confidential_remarks":"The central obstacle to acceptance is the state-dimension reduction in Section 5: the experiments appear to solve a two-dimensional surrogate problem, while the abstract claims generalization across system dynamics. If the authors can validate against full-order ground-truth solutions or sharply reframe the claims to the reduced problem, the paper would be publishable. I also note that the manuscript does not mention availability of code or data; making the data-generation pipeline and model training code available would substantially strengthen the reproducibility of the work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is the first paper I've seen that frames BRT computation as operator learning between value function spaces, and it shows that off-the-shelf FNO/TNO can map signed-distance functions to converged reachability value functions with held-out relative L2 errors in the 0.002–0.037 range and inference times around a millisecond. Second, the main claim of generalization across system dynamics is not supported, because the paper treats velocity and heading as constant hyperparameters rather than evolving states.\n\nThe good parts: the problem framing is clean, the training data comes from an external HJ solver, test sets are held-out obstacle shapes, and the zero-shot super-resolution check is a nice touch. The two-obstacle experiment correctly notes that the union of single-obstacle solutions is not the correct BRT. The parametric-inputs experiment, which varies acceleration and curvature limits, is legitimate because those are genuine scalars.\n\nThe soft spot is load-bearing. Section 5 says V∞ 'typically varies only with the spatial coordinates... while remaining approximately constant along dimensions such as velocity or heading,' and then encodes v and θ as constant inputs via Eq. (26). But for the dynamic Dubins car, v and θ are states with dynamics v˙=u1 and θ˙=v u2. A 2D slice at fixed v,θ is the solution of a different, lower-dimensional system; it is not the value function of the full-order system. The velocity-dependent experiment makes this concrete: it deliberately makes the obstacle radius grow with v, so V∞ varies with v by construction, and encoding v as a constant input cannot produce a 4D value function. No comparison against full-order ground truth is given. The only full-state result (Air3D) shows degradation (FNO 0.028→0.050, TNO 0.037→0.15), which is consistent with the full problem being qualitatively harder rather than validating the reduced model. I'd also like error bars, baselines, and code, but those are minor by comparison.\n\nNone of this is circular: training data come from an external solver, and the core empirical result is plausible for the reduced 2D setup. But the abstract's statement about generalizing 'across varying system dynamics' goes beyond what is shown.\n\nWho should read it: anyone working on learning-based reachability or fast safety filters for low-dimensional systems with fixed dynamics. It deserves a serious referee, but it needs substantial revision: either validate the reduction against full-order ground truth for at least a few cases, or re-frame the contribution as learning on a reduced model with the reduction stated as an assumption rather than a fact.","headline":"First neural-operator treatment of BRTs with genuinely fast inference, but the reduction of velocity and heading to constant hyperparameters means the reported errors may be for a surrogate 2D problem, not the true dynamics.","tokens_in":10723,"tokens_out":2624,"would_cite":false,"duration_ms":24260,"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 neural operator trained once can map an obstacle geometry directly to the safety region of an autonomous system, predicting unseen scenarios in milliseconds with relative L2 errors between 0.002 and 0.037.","keywords":["Hamilton-Jacobi reachability","neural operators","Fourier Neural Operator","backward reachable tube","safety analysis","autonomous systems","value function","generalization"],"falsifier":"Take a Dubins car with a wide speed range and an obstacle whose clearance grows rapidly with speed (e.g., radius as an exponential of $v$, as in the velocity-dependent experiment), solve the full 4D HJI equation with a grid solver, and compare the predicted $V_\\infty$ from the 2D-reduced HJRNO model to the full-state ground truth; if the relative L2 error is far above the reported 0.002-0.037 band, the central generalization claim fails for genuinely velocity-dependent dynamics.","tokens_in":9652,"feed_emoji":"🚗","tokens_out":6628,"duration_ms":58041,"temperature":0.7,"pith_summary":"The paper claims that the computationally expensive Hamilton-Jacobi-Isaacs (HJI) PDE solution behind backward reachable tubes (BRTs) can be replaced by a neural operator that maps the initial value function -- equivalently the obstacle geometry -- directly to the converged value function $V_\\infty(x)$. Training once on randomly generated obstacle scenarios suffices to predict unseen scenarios with relative L2 errors between 0.002 and 0.037 and inference times near 0.001 to 0.005 seconds, orders of magnitude faster than traditional grid solvers. If true, this makes real-time safety analysis for autonomous systems practical: safety filters could be updated as the environment changes without re-solving the PDE. The main reduction treats non-spatial states such as velocity and heading as constant hyperparameters, which keeps the model two-dimensional.","feed_headline":"Safety reachability computed in 0.001 seconds with a neural operator","feed_subtitle":"Trained once, it predicts unseen obstacle scenarios with relative errors under 4 percent.","key_machinery":"The central mechanism is the learned operator equation $s = \\mathcal{M}_\\theta(a)$, instantiated with two neural operator architectures: the Fourier Neural Operator, which performs global convolution in the spectral domain using the convolution theorem, and the Galerkin Transformer, which approximates the integral operator via self-attention. Both map the input function $a(x)=V(t=0,x,h)$ sampled on a uniform grid to the output $V_\\infty(x,h)$ and are discretization-invariant, so the same model can be evaluated at new resolutions. The non-spatial state dimensions are folded into the input as constant functions $\\tilde h(x)=h$, reducing the effective problem to two spatial coordinates and making the mapping tractable for FNO and TNO.","core_discovery":"On its own terms, the discovery is that the operator $\\mathcal{M}_\\theta$ taking the initial value function $a(x)=V(t=0,x,h)$ to the converged value function $s(x)=V_\\infty(x,h)$ is learnable by neural operators, specifically the Fourier Neural Operator and the Galerkin Transformer. With scalar system hyperparameters embedded as constant functions over the state domain, a single trained model generalizes across random obstacle shapes, multi-obstacle layouts, indoor environments, velocity-dependent obstacle sizes, and continuous control-limit variations. Across six experimental setups the relative L2 test error stays between 0.002 and 0.037, inference takes about 0.001 to 0.005 seconds, and the trained model also shows zero-shot super-resolution, producing accurate predictions at resolutions higher than trained. The paper positions this as the first application of neural operators to Hamilton-Jacobi reachability, eliminating the need to retrain the PDE solver for each new problem setting.","pith_inferences":["If the mapping truly remains accurate on systems where velocity or heading materially influence the value function, the constant-hyperparameter reduction could be dropped and the same operator framework trained directly on higher-dimensional states; the paper reports degradation on full-state Air3D, so this is an open pressure point rather than a demonstrated claim.","The parametric-input trick suggests a natural extension: add time as an input dimension to learn finite-horizon reachable tubes $V(t,x)$ in one model, rather than only the converged $V_\\infty$.","Because the training data come from a grid solver, HJRNO inherits any systematic bias of that solver; the reported errors are with respect to the same solver's solutions, so end-to-end safety guarantees would require validating against an independent solver or an exact solution.","A testable industrial extension: deploy HJRNO in a simulated warehouse or road scenario with dynamic obstacles and measure whether the safety-filter decisions differ from ground-truth BRT decisions in ways that matter for collision avoidance."],"forward_implications":["Safety filters for autonomous vehicles can be updated in real time as obstacles move or change shape, because inference is orders of magnitude faster than solving the HJI PDE.","A single trained model covers a family of problem settings -- obstacle geometries and system hyperparameters -- so no instance-specific retraining is needed.","Because neural operators are discretization-invariant, models trained at lower resolution can be deployed at higher resolution without retraining (zero-shot super-resolution).","The learned operator can serve as a fast surrogate inside safety-critical loops on onboard computers, with checkpoint sizes (3-38 MB) comparable to storing one solution instance of a 4D state.","Arbitrary and non-convex domains are within reach because neural operators extend to general geometries, not just uniform grids."],"supporting_citations":[{"why":"Supplies the Fourier Neural Operator architecture used to learn the value-function mapping.","marker":"[18]"},{"why":"Supplies the Galerkin Transformer architecture (TNO) used as the second neural operator.","marker":"[7]"},{"why":"The JAX-based Hamilton-Jacobi reachability solver used to generate all training and test value-function pairs.","marker":"[21]"},{"why":"Provides the theoretical grounding for neural operators as discretization-invariant universal approximators, justifying the function-space mapping approach.","marker":"[16]"},{"why":"Defines the Hamilton-Jacobi reachability problem and backward reachable tubes that the paper solves.","marker":"[2]"},{"why":"The prior parameter-conditioned reachability approach that only handles scalar hyperparameters, which HJRNO extends to functional inputs.","marker":"[4]"}],"fun_headline_variants":["One neural operator handles all obstacle shapes for safety reachability","Neural operator predicts safety tubes in 0.001s across unseen scenarios","First neural operator for Hamilton-Jacobi reachability: fast and general","Safety analysis goes real-time with neural operator foundation model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the value function is effectively two-dimensional: velocity and heading are treated as constant hyperparameters rather than states, so systems where the value function varies strongly along those dimensions are not actually represented by the learned mapping.","fun_headline_variants_meta":{"raw":{"variants":["One neural operator handles all obstacle shapes for safety reachability","Neural operator predicts safety tubes in 0.001s across unseen scenarios","First neural operator for Hamilton-Jacobi reachability: fast and general","Safety analysis goes real-time with neural operator foundation model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1405,"prompt_tokens":852,"completion_tokens":553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":481}},"tokens_in":468,"tokens_out":553,"duration_ms":5490,"temperature":1.0,"reasoning_tokens":481,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:37:53.158717+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a Dubins car with a wide speed range and an obstacle whose clearance grows rapidly with speed (e.g., radius as an exponential of $v$, as in the velocity-dependent experiment), solve the full 4D HJI equation with a grid solver, and compare the predicted $V_\\infty$ from the 2D-reduced HJRNO model to the full-state ground truth; if the relative L2 error is far above the reported 0.002-0.037 band, the central generalization claim fails for genuinely velocity-dependent dynamics.","supporting_citations":[{"cited_title":"hj reachability: Hamilton-jacobi reachability analysis in jax","cited_arxiv_id":null,"evidence_quote":"The JAX-based Hamilton-Jacobi reachability solver used to generate all training and test value-function pairs."},{"cited_title":"Neu- ral operator: Learning maps between function spaces with applications to pdes","cited_arxiv_id":null,"evidence_quote":"Provides the theoretical grounding for neural operators as discretization-invariant universal approximators, justifying the function-space mapping approach."},{"cited_title":"Hamilton- jacobi reachability: A brief overview and recent advances","cited_arxiv_id":null,"evidence_quote":"Defines the Hamilton-Jacobi reachability problem and backward reachable tubes that the paper solves."},{"cited_title":"Parameter- conditioned reachable sets for updating safety assurances online","cited_arxiv_id":null,"evidence_quote":"The prior parameter-conditioned reachability approach that only handles scalar hyperparameters, which HJRNO extends to functional inputs."}],"review_version":1}