{"id":"a1ba4b8f-f1d5-4b4b-8fe8-3ddd296ceb72","arxiv_id":"2505.03830","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MPC-generated approximate value labels guide a DeepReach-style network to learn Hamilton-Jacobi reachability solutions, yielding larger verified safe sets in 2D, 7D, 13D, and 40D systems.","lead":"This paper trains neural safety value functions for high-dimensional robots by adding model predictive control labels to the usual Hamilton-Jacobi PDE loss. On drone, race-car, and 40D benchmarks, the hybrid method recovers larger verified safe sets than existing learning-based reachability baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Iterative MPC-label refinement can lock in the network's own bias; no experiment or analysis tests whether regenerated labels are closer to ground truth.","rationale":"The paper's central claim is that MPC-generated approximate values, used as semi-supervised labels and refined during training, yield more accurate and stable HJB-VI solutions. For that claim, the refinement step must make labels better, not merely self-consistent. Algorithm 2 refines DMPC exactly when the network is being used to define the terminal cost; the same network was trained on the previous DMPC, so the loop is circular in the absence of an error-contraction argument. The paper's only ablation (Sec. IV-D) shows that turning refinement off hurts on the 2D drone, but this does not establish that refinement is safe: a single seed, no ground-truth comparison of the labels themselves, and no test of how the loop responds to a biased initialization. I agree with the reader's weakest assumption. The proposed test, instrumenting the refinement rounds against OptimizedDP ground truth on the vertical drone and repeating from a biased warm start, would directly determine whether errors decay or propagate. Other weaknesses, such as the fixed-horizon labeling ambiguity in Algorithm 1 versus Eq. (8), the >100% F1Tenth volume in Table II, and single-seed headline results, further support a conditional verdict, but the refinement loop is the most load-bearing because it is the novel mechanism that the paper claims mitigates local optima. Verdict remains CONDITIONAL.","tokens_in":16303,"tokens_out":11103,"duration_ms":126720,"concrete_test":"On the vertical-drone benchmark (ground truth from OptimizedDP), instrument Algorithm 2: at every refinement round r, record (a) the L2 error of V_theta at the terminal states used as terminal cost in the MPC rollouts and (b) the L2 error of the newly generated DMPC labels at the corresponding initial states, both versus ground truth, along with the sign of the errors. If error (b) is not consistently smaller than error (a), or if the errors share the same sign (both over-optimistic), the refinement is propagating the network's bias rather than correcting the MPC approximation. A second run from a deliberately biased warm start (e.g., V_theta initialized to the Vanilla DeepReach solution) and comparison of the final recovered volume against the standard initialization would show whether the loop is self-correcting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the dataset-refinement loop in Algorithm 2 (lines 12-14). When the curriculum time t crosses t_R, the MPC labels are regenerated using the current network V_theta as terminal cost over [t_R-H_R, t_R]. For this to improve labels, V_theta must already be closer to the true value function than the MPC approximation it replaces. The paper provides no mechanism or bound ensuring this. Because V_theta was itself trained partly on the previous (biased) MPC labels, the new labels can inherit and amplify the same bias: if V_theta is over-optimistic on the boundary at t_R, the terminal cost is too high, the new labels are too optimistic, and the supervised term Ldata = ||V_hat - V_theta|| then anchors the network to its own earlier mistake. The fine-tuning loss (12) only upweights false positives after curriculum training; it does not correct a systematically over-optimistic label set. The only direct ablation of refinement (Sec. IV-D) is a single 2D vertical-drone run (MSE 0.2285 vs 0.009, volume 20.04% vs 24.62%) with no seeds, no error bars, and no measurement of whether the regenerated labels were actually closer to the ground truth. A system whose labels are generated from the current solution and whose loss enforces agreement with those labels is at risk of self-confirmation; the paper neither analyzes this feedback nor tests it adversarially.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid approach to learning Hamilton-Jacobi reachability value functions in which a sampling-based MPC routine (Algorithm 1) generates approximate value-function labels at collocation points, and a neural network is trained with a weighted combination of the HJB-VI PDE residual (Eq. 10) and a data-supervision term (Eq. 9). Training proceeds through pretraining, curriculum training, and fine-tuning (Algorithm 2), with the MPC dataset periodically regenerated using the current learned value function as terminal cost. The learned value function is then processed through a conformal-prediction verification step to produce a high-confidence safe set. Experiments on four systems (2D vertical drone, 13D quadrotor, 7D F1Tenth, 40D publisher-subscriber) are reported, comparing against vanilla DeepReach, distillation, and neural CBF baselines, with an additional ablation of dataset size and training stages.","tokens_in":16620,"tokens_out":8342,"duration_ms":77325,"significance":"If validated, the core idea is valuable: it injects a strong optimal-control-based learning signal into physics-informed reachability, potentially mitigating the weak-supervision problem that causes residual-only training to converge to nonphysical solutions. The paper includes an open-source implementation and demonstrates large reported gains on high-dimensional examples (e.g., 93.69% vs 71.26% recovered volume on the 13D quadrotor). The conformal verification step is a useful practical addition, as it connects the learned value function to a probabilistic safety guarantee. However, the current validation has important gaps: single-seed main results, a missing benchmark against the closest label-based prior, an unexplained 167.14% volume for F1Tenth Neural CBF, and no analysis of the dataset-refinement feedback loop. The time-indexing inconsistency in Algorithm 1 is a serious technical issue that must be resolved before the method can be reproduced.","major_comments":[{"comment":"The sampled time t_i is not used in the rollout, so the stored label V̂(t_i,x_i) is computed from a fixed-horizon H rollout starting at x_i at time 0, not from horizon T−t_i. Since the value function depends on the remaining horizon, the label is inconsistent with the stated (x,t) unless HΔ=T and t_i=0. The same issue occurs in Remark 2, where bootstrapped labels are assigned time Δ·h instead of t_i+Δ·h. Please clarify how t_i enters the label generation (e.g., set H per sample as (T−t_i)/Δ) or correct the time indexing; as written, Eq. (9) trains against labels that may not be evaluations of the target value function at the reported times.","section":"Section III-A, Algorithm 1, Remark 2"},{"comment":"The dataset-refinement loop uses the current learned V_theta as terminal cost for regenerating D_MPC, with no mechanism or bound ensuring that the new labels are closer to the true value function than the labels they replace. If V_theta is over-optimistic, the refined labels inherit that bias and the supervised loss in Eq. (9) can reinforce it. The only ablation is a single vertical-drone run (MSE 0.2285 vs 0.009, volume 20.04% vs 24.62%) with no seeds or measurement of label quality. To support the claim that refinement 'mitigates convergence to local optima,' please provide multi-seed ablations on more than one system and, where ground truth exists, report the error of the regenerated labels versus the ground truth across refinement iterations.","section":"Algorithm 2, Section IV-D"},{"comment":"The F1Tenth entry for Neural CBF reports a recovered volume of 167.14%, which exceeds 100% of the state-space volume and is not a valid volume fraction. The table also reports single-run values for all methods, so the claim that the proposed approach 'consistently' achieves higher volumes is not backed by variance estimates. Please correct the metric (or explain what is being measured) and report mean and standard deviation over at least five seeds for the main comparisons, as is already done for the vertical-drone dataset-size ablation.","section":"Table II, Sections IV-B and IV-F"},{"comment":"The introduction identifies the Hopf-formula semi-supervised method [32] as the closest label-based alternative to the proposed approach, but [32] is not included in the benchmark. Without a comparison to this method, the claim that the MPC-guided approach outperforms 'existing methods' is incompletely supported. Please add this baseline or justify its omission.","section":"Sections I and IV-A"},{"comment":"The false-positive fine-tuning loss is written as λ_FP∥V̂−Vθ(x,t)∥Vθ(x,t) when Vθ(x,t)≥0 and V̂<0, while the text states that the loss is amplified by the scalar λ_FP. The extra multiplicative Vθ(x,t) makes the amplification vanish at Vθ=0 and grow with the value magnitude, which is not what the text describes. Clarify whether this factor is intended, and if not, correct the equation, since the exact loss is essential for reproducibility.","section":"Section III-B, Eq. (12)"}],"minor_comments":[{"comment":"The word 'exasperated' should be 'exacerbated'.","section":"Section IV-F"},{"comment":"The parenthetical entries for the vertical drone (31.28 and 56.51) are not explained in the table caption; add a note defining that these correspond to K=12.","section":"Table II caption"},{"comment":"References [32] and [33] point to the same arXiv paper; merge to avoid duplicate citation.","section":"References"},{"comment":"The confidence value β=10^{-16} is extremely small; please report the number of calibration samples used in the conformal procedure so the reader can judge the statistical validity.","section":"Section III-C"},{"comment":"The PDE collocation times are drawn uniformly in [t,T], but the boundary at t=T is explicitly enforced by the ansatz; it may be helpful to state this in the text.","section":"Algorithm 2, line 7"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is promising but needs substantial revision. The most serious issue is the time-index inconsistency in Algorithm 1, which may indicate that the method as implemented differs from the written algorithm. If that issue is resolved, the remaining concerns (refinement circularity, missing baseline, statistical reporting) are addressable with additional experiments. I would not recommend rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a practical win. Using sampling-based MPC to generate approximate value function labels and mixing them with PDE residual loss makes neural HJ reachability training noticeably more robust and accurate on hard cases. The 13D quadrotor and 40D publisher-subscriber results are genuinely impressive, and the conformal prediction verification gives the safe sets a meaningful guarantee.\n\nWhat is genuinely new is the specific combination: MPC label generation with bootstrap-style running values, a temporal curriculum, and iterative dataset refinement. That combination is not in the DeepReach, PINN, or Hopf-supervision literature, and the MPC Distillation baseline is a clean way to show that labels alone are not enough. The paper also ships code, which is good.\n\nThe soft spots are real but not fatal. The most serious is the refinement loop in Algorithm 2. When the dataset is regenerated using the current V_theta as terminal cost, labels can inherit the network's own bias, and the supervised loss can then anchor the network to that same bias. The paper's only ablation of refinement is a single 2D vertical drone run with no seeds and no measurement of whether regenerated labels are actually closer to ground truth. That is a weak spot, and I'd want either a theoretical bound or a targeted experiment before fully trusting the loop.\n\nAlso missing is a benchmark against the Hopf-label method [32], which is the most relevant label-based prior art. The main-table results are single-seed; the five-seed ablation only covers dataset size on the drone. And Table II shows Neural CBF recovering 167% volume on F1Tenth, which is physically impossible — the paper notes it but never explains the number. Those are fixable issues.\n\nThe central claim still holds up: MPC guidance does help, and the improvement over Vanilla DeepReach is consistent across all four case studies. The refinement concern is a genuine risk, but nothing in the paper demonstrates that it is catastrophic in practice.\n\nThis paper deserves a serious referee. I would send it out and ask for the Hopf baseline, multi-seed stats for the main comparisons, and an explicit analysis of the refinement loop. It is a useful contribution for the reachability and safe-control community.","headline":"Solid engineering contribution showing MPC-generated value labels improve neural HJB reachability learning, but the iterative refinement loop deserves deeper scrutiny.","tokens_in":17135,"tokens_out":1524,"would_cite":true,"duration_ms":16556,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MPC-generated value labels steer neural reachability training to accurate, verifiable safe sets.","keywords":["Hamilton-Jacobi reachability","backward reachable tube","model predictive control","physics-informed neural networks","HJB variational inequality","neural value function","conformal prediction","safety verification"],"falsifier":"Take a low-dimensional system with a known ground-truth value function and a long-horizon avoidance problem where random-shooting MPC systematically misses the optimal trajectory (for example, a U-shaped obstacle requiring a detour longer than the sampling horizon). Train the proposed method with a deliberately biased initial value function (e.g., pre-trained to label all states safe) and start refinement immediately. If the recovered verified volume stays at or below the residual-only DeepReach baseline, the central claim that MPC labels guide the network to the true HJB-VI solution is falsified.","tokens_in":16073,"feed_emoji":"🤖","tokens_out":5265,"duration_ms":53439,"temperature":0.7,"pith_summary":"Hamilton–Jacobi (HJ) reachability gives a safety value function whose sub-zero level sets are the states from which failure is inevitable, but grid solvers scale exponentially with dimension and residual-only neural training often locks onto non-physical solutions. This paper tries to establish that approximate value labels generated cheaply by sampling-based model predictive control (MPC) — solving the same optimal control problem that underlies HJ reachability — can guide a neural network to the correct HJB variational inequality solution. The authors add these labels as a supervised loss alongside the PDE residual, re-fit the MPC dataset during training using the current value function as terminal cost, and verify the result with conformal prediction. On a 2D drone, a 13D quadrotor, a 7D F1Tenth car, and a 40D publisher-subscriber system, the claimed outcome is a substantially larger verified safe set than residual-only or distillation-only training. The payoff would be practical, verified reachability analysis for higher-dimensional nonlinear systems.","feed_headline":"MPC labels let neural reachability recover 97% of the true safe set","feed_subtitle":"Residual-only training collapses; MPC-guided training verified 76-97% on four robot systems.","key_machinery":"The machinery is a MPC dataset generator (Algorithm 1) plus a three-phase training loop (Algorithm 2). Algorithm 1 samples control sequences around a nominal policy, rolls out discretized dynamics, and records the best worst-case safety cost $\\hat V(t_i,x_i)=\\max_u \\min_{h} l(\\xi_h)$, bootstrapping labels along the best trajectory; it is GPU-parallel and treats the learned $V_\\theta$ as terminal cost when the horizon extends. Algorithm 2 then pretrains on these labels, performs curriculum training with combined loss $L = L_{\\mathrm{PDE}} + \\lambda L_{\\mathrm{data}}$ and periodic dataset refinement every $H_R$ seconds, and fine-tunes with an asymmetric false-positive penalty. This machinery is what carries the paper's claim that approximate optimal-control data can steer a residual-trained network away from bad local solutions.","core_discovery":"Starting from the observation that the backward reachable tube is the solution of the optimal control problem $V(x,t) = \\sup_{u(\\cdot)} \\min_{\\tau\\in[t,T]} l(\\xi(\\tau))$, the paper claims that a sampling-based MPC solver computing $\\hat V$ at collocation points $(t_i,x_i)$ produces valuable semi-supervised labels for learning the HJB variational inequality $\\min\\{D_t V + H, l - V\\}=0$ with terminal condition $V(x,T)=l(x)$. Training proceeds in three phases: supervised warm-start on the MPC dataset, curriculum training that jointly minimizes the PDE residual and a data-matching loss while periodically refining the MPC dataset with the current learned value function as terminal cost, and a fine-tuning phase that up-weights false-positive errors to keep the safe set conservative. The paper claims this hybrid training is more accurate and stable than residual-only DeepReach and more accurate than pure MPC distillation, with verified safe-set volumes of 24.62%, 93.69%, 76.08%, and 97.14% on the four case studies.","pith_inferences":["A natural reading is that the approach inherits the failure profile of the MPC guide: wherever random-shooting MPC cannot approximate the optimal avoidance (long detours, narrow passages), the supervised loss could pull the network toward a suboptimal value function, which the paper's ablations do not isolate.","Because the method's cost scales with problem complexity rather than nominal dimension (7D F1Tenth costs more than 13D quadrotor), a promising extension is adaptive sampling of collocation points informed by PDE residual or label disagreement to cut MPC calls further.","The refinement loop suggests a self-consistency criterion: labels and residuals should agree as training converges, so monitoring the gap between $L_{\\mathrm{data}}$ and $L_{\\mathrm{PDE}}$ could detect when the MPC guides have gone stale or biased.","The same MPC-guidance scheme could in principle be ported to reach-avoid problems or disturbance-robust games by replacing the single-player maximization with a saddle-point robust MPC, though the paper explicitly leaves disturbances and long-horizon problems out."],"forward_implications":["If the central claim is right, learned safe sets can be verified: the recovered verified volumes (24.62% on the vertical drone, 93.69% on the 13D quadrotor, 76.08% on F1Tenth, 97.14% on the 40D system) exceed every baseline tested.","Pure MPC distillation is insufficient: the same MPC labels without HJB residuals yield zero verified volume on both the drone and F1Tenth, so the PDE residual remains necessary.","The time curriculum is the load-bearing training stage: removing it collapses recovered volume to zero on the quadrotor and F1Tenth, with negligible cost savings.","Iterative dataset refinement is directly beneficial: on the vertical drone, keeping the MPC dataset fixed drops recovered volume from 24.62% to 20.04% and raises MSE to 0.2285.","The learned safety policy is usable for downstream filtering: on F1Tenth it turns a nominal policy with 100% collision rate and 36.17 m average travel into a collision-free filtered policy averaging 187.01 m."],"supporting_citations":[{"why":"Provides the DeepReach paradigm this paper builds on: learning the HJB reachability solution by minimizing PDE residuals with a neural network.","marker":"[2]"},{"why":"Is the Vanilla DeepReach baseline with exact boundary condition imposition, against which the proposed method is compared.","marker":"[34]"},{"why":"Supplies the conformal prediction verification scheme that converts the learned value function into a probabilistic safe set.","marker":"[22]"},{"why":"Supplies the OptimizedDP ground-truth value function used to compute MSE and false-positive rates on the low-dimensional drone case.","marker":"[4]"},{"why":"Gives the 40D publisher-subscriber system and its decomposition-based ground truth, enabling the high-dimensional accuracy comparison.","marker":"[33]"},{"why":"Motivates the added data-driven loss term with the analogy to physics-informed neural networks that anchor trial solutions at collocation points.","marker":"[5]"}],"fun_headline_variants":["MPC-guided training recovers 97% of safe set in neural reachability","Verified safe sets up to 97% with MPC-guided neural reachability","MPC labels fix deep reachability training, hitting 97% accuracy","Residual-only training collapses; MPC guidance hits 97% safe sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The refinement loop assumes the current learned value function is already accurate enough to serve as the terminal cost and nominal policy when the MPC dataset is regenerated; if early curriculum labels are biased, the new MPC labels inherit that bias and the data loss can lock the network into the wrong reachable set.","fun_headline_variants_meta":{"raw":{"variants":["MPC-guided training recovers 97% of safe set in neural reachability","Verified safe sets up to 97% with MPC-guided neural reachability","MPC labels fix deep reachability training, hitting 97% accuracy","Residual-only training collapses; MPC guidance hits 97% safe sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000531,"raw_usage":{"total_tokens":2595,"prompt_tokens":1022,"completion_tokens":1573,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":1492}},"tokens_in":638,"tokens_out":1573,"duration_ms":11725,"temperature":1.0,"reasoning_tokens":1492,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:04:40.450513+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a low-dimensional system with a known ground-truth value function and a long-horizon avoidance problem where random-shooting MPC systematically misses the optimal trajectory (for example, a U-shaped obstacle requiring a detour longer than the sampling horizon). Train the proposed method with a deliberately biased initial value function (e.g., pre-trained to label all states safe) and start refinement immediately. If the recovered verified volume stays at or below the residual-only DeepReach baseline, the central claim that MPC labels guide the network to the true HJB-VI solution is falsified.","supporting_citations":[],"review_version":1}