{"id":"1341f32b-2161-461c-a892-53697f378e1c","arxiv_id":"2505.05691","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A self-supervised neural planner combines Eikonal, temporal difference, obstacle alignment, and causality losses with a learned L1 and L-infinity metric, improving success rates and generalization in robot motion planning.","lead":"This paper trains a neural network to solve the Eikonal equation, a mathematical model of travel time, for robot motion planning by adding temporal difference and metric learning losses. It reports higher success rates and faster planning than prior self-supervised methods in simulated mazes, cluttered 3D spaces, and 7 to 12 degree of freedom robot arms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 6's obstacle-normal term is ill-defined exactly where it acts: after the Eq. 2 clip, ∇S* vanishes for dobs<dmin while (1−S*)>0, leaving ∇S*/||∇S*|| as 0/0.","rationale":"The reader's weakest assumption identified non-smoothness of the ground-truth speed and its gradients. My concern is the same underlying dependency on ∇S*, but more precise: even with a smooth dobs, the clipping in Eq. 2 creates saturated regions where ∇S* is exactly zero and the Eq. 6 normal term is undefined. This is a concrete mathematical defect in a method component whose ablation shows it contributes to the reported performance. It does not by itself disprove the central empirical claim, since a working implementation presumably uses some surrogate, but it makes the published method non-reproducible as stated and shifts the burden onto an undocumented implementation choice. I do not see a reason to move the verdict to REJECT: the TD/Eikonal derivation is coherent, the comparisons are plausible, and the empirical trends are consistent. The issue is exactly the kind that should be resolved before acceptance, so the existing CONDITIONAL verdict is appropriate and unchanged. I marked agreement as partial because the reader pointed to general differentiability of dobs, while the load-bearing failure is the saturation of the clipped S*, which remains even under the reader's assumed differentiability.","tokens_in":18239,"tokens_out":16938,"duration_ms":214712,"concrete_test":"Re-run the maze ablation of Fig. 3 / Table 2 with the exact published loss: compute S* by Eq. 2, autodiff ∇S* through the clip, and evaluate Eq. 6 at 10^4 sampled configurations with dobs < dmin. Check whether the term is NaN, zero, or undefined. If it is, the formula as printed cannot be what was trained; then retrain with a soft-clipped or unclipped-gradient variant of S* and see whether the reported 0.08 error and the −LN ablation behavior are preserved. This directly settles whether Eq. 6 is well-defined and whether the claimed LN contribution is reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The obstacle-normal alignment loss in Eq. 6 is a central part of the proposed loss combination, and the ablation in Table 2 shows it matters (removing LN raises the maze error from 0.08 to 0.13). But as written, Eq. 6 cannot be evaluated in the region where it is supposed to be active. Eq. 2 defines S* = clip(dobs/dmax, dmin/dmax, 1). Consequently, for every configuration with dobs(q) < dmin, S* is identically dmin/dmax and ∇S* = 0; in that same region the weight (1−S*) is positive and typically close to 1. The term ∇S*/||∇S*|| is therefore 0/0 precisely where the loss has its largest influence. This is not a measure-zero artifact: the saturated near-obstacle shell has positive measure, and random training samples will land there. The paper's assumption in Sec. 3.2 that dobs is differentiable does not resolve the issue, because even a perfectly smooth dobs becomes constant after clipping. The far-field case (S* = 1) is also non-differentiable at the clip boundary, though the weight 1−S* makes that term formally zero. The paper never states a surrogate such as using the unclipped distance gradient in the normal term or clamping the denominator. So the actual trained loss is underspecified: either Eq. 6 is undefined on a substantial part of the training distribution, or an undocumented modification is needed to make it finite. Since the central claim rests on the specific loss combination, this is a load-bearing reproducibility and correctness concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a self-supervised method for robot motion planning that learns a travel-time function T(qs,qg) by solving the Eikonal equation. The training objective in Eq. (8) combines the standard Eikonal loss, a temporal difference loss enforcing Bellman optimality over a finite step, an obstacle-normal alignment loss, and a causality weight. The travel-time function is parameterized as a metric distance in a learned latent space using a blockwise L1/L-infinity norm, and the encoder is conditioned on environment point clouds through attention, which enables generalization to unseen environments. Planning at test time is performed with sampling-based MPC. Experiments on 2D mazes, Gibson indoor scenes, C3D cluttered 3D scenes, 7-DOF Franka manipulation, and a 12-DOF dual-arm real-world cabinet show improved success rates and lower planning times relative to prior self-supervised planners.","tokens_in":18703,"tokens_out":14860,"duration_ms":150179,"significance":"The paper makes a solid empirical case that adding a finite-horizon Bellman consistency term to the Eikonal loss improves the quality of the learned travel-time field, and that a metric-learning architecture with a non-Euclidean latent distance is beneficial. The ablation study in Table 2 quantifies the contribution of each loss component and of the metric choice, and the reported success rates on multi-DOF tasks are strong. The environment-conditioned attention mechanism is a useful step toward generalization to unseen scenes, and the promise of code release is a positive reproducibility signal. However, the current manuscript contains an undefined loss term in a load-bearing component, and the empirical gains are not fully disentangled from the MPC inference scheme, so the central claim of a more accurate Eikonal solution is not yet fully supported.","major_comments":[{"comment":"The obstacle-normal alignment loss L_N is ill-defined on a set of positive measure. Since S* is defined by the clipping in Eq. (2), for any configuration with d_obs(q) < d_min, S* is constant at d_min/d_max and hence ∇S* = 0, while the weight (1−S*) is positive. The expression ∇S*/||∇S*|| is therefore 0/0 precisely in the near-obstacle region where the loss is intended to be active. The differentiability assumption on d_obs in §3.2 does not resolve this issue because clipping eliminates the gradient. The manuscript does not specify a surrogate, such as using the gradient of the unclipped distance or clamping the denominator, so the actual trained objective is underspecified. Since the ablation in Table 2 shows that removing L_N raises the maze error from 0.08 to 0.13, this is a load-bearing reproducibility and correctness concern.","section":"§4.1.2, Eq. (6)"},{"comment":"The Ours-G variant, which uses gradient-based path inference instead of MPC, is reported only for the Gibson environment in Table 1(a). In the C3D and 7-DOF manipulator experiments, the comparison of Ours (MPC) against NTF and P-NTF (gradient-based inference) is confounded by the inference mechanism. It is therefore possible that the higher success rates in Tables 1(b) and 1(c) come from MPC's stochastic exploration rather than from a more accurate learned value function. The authors should report Ours-G results for these tasks or explicitly disentangle the contribution of MPC from that of the learned Eikonal solution to support the central claim that the proposed losses and architecture improve the value function.","section":"§5.2, Tables 1(b) and 1(c)"}],"minor_comments":[{"comment":"The causality weight L_C = exp(-λ_C T(qs,qg)) multiplies all losses, including the Eikonal loss. While this is a known heuristic from Wang et al. (2024b), the paper does not analyze whether this output-dependent weighting changes the set of stationary points of the combined loss. Even though the true solution is still a stationary point, spurious minima may be introduced; a brief empirical or theoretical note would strengthen the claim that causality preservation is one of the three key enhancements.","section":"§4.1.3, Eqs. (7)-(8)"},{"comment":"The claim that the L1/L∞ metric 'preserves the geodesic structure' and supports multiple shortest paths is supported only by the illustration in Fig. 2. Please clarify whether this is a rigorous property of the proposed construction or an empirical observation, and provide a mathematical statement if available.","section":"§4.2.1"},{"comment":"The hyperparameters d_min, d_max, a, and b in Eqs. (2) and (10) are never specified. Since these define the ground-truth speed and the latent space geometry, please include their values in Appendix C or in the code release to make the experiments reproducible.","section":"§3.2 and §4.2.1"},{"comment":"The hyperparameters λ_E, λ_TD, λ_N, λ_C, and Δt are given, but the dependence of Δt and λ_N on the environment is described only qualitatively; please provide the specific values used for each environment.","section":"Appendix C.2"},{"comment":"There are several typos and reference inconsistencies: 'Our-G' vs 'Ours-G' in §5.2; 'metirc' in §4.2.1; 'Demontration' in the Fig. 5 caption; and the P-NTF citation in §5 lists (Ni & Qureshi, 2023a) while the related work and §3.2 cite (Ni & Qureshi, 2023b).","section":"§5.2 and figure captions"},{"comment":"The caption begins with 'L1L2', which appears to be a leftover label; the figure should be clarified with respect to the L1 and L∞ components of the proposed metric.","section":"Fig. 2 caption"},{"comment":"The sampling-based MPC is described qualitatively, but the number of samples, horizon, and softmax temperature are not reported; please include these parameters in Appendix C.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong empirical contribution, but the ill-defined L_N term in Eq. (6) and the incomplete Ours-G ablation across tasks are substantive issues that need to be resolved before the central claims can be accepted. I recommend major revision rather than rejection, as both issues appear addressable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the takeaway: the paper's obstacle-normal alignment loss (Eq. 6) is undefined in the region where it matters, because of the clipping in Eq. 2. The stress-test note is correct. That's a real flaw, not a quibble. But the rest of the paper is better than the flaw, and the empirical story deserves scrutiny rather than a desk reject.\n\nWhat's new: the authors frame the Eikonal solution as an optimal value function and a geodesic, then add a temporal difference loss to the usual Eikonal loss, plus a metric-learning architecture (blockwise L1/L∞) that respects triangle inequality and symmetry. They also condition on the environment point cloud with attention and use sampling-based MPC at inference. Each piece is motivated and ablated. The maze ablation shows every loss component helps, and the metric comparison against IQE, PQE, MRN, DN is fair. The main results are strong: 99.4% success on C3D, 91% on a 12-DOF dual-arm, with about 5-10x speedup over classical baselines.\n\nThe soft spots, in proportion. First, Eq. 6 is broken as written. S* is clipped, so inside the near-obstacle shell dobs<dmin, the gradient ∇S* is zero while the weight (1-S*) is near one. The term ∇S*/||∇S*|| becomes 0/0 exactly where the loss is active. The paper never defines a surrogate. Since the ablation shows removing LN hurts, something is being computed, but it's not shown. This must be fixed before trusting the method.\n\nSecond, the gradient-inference variant Ours-G only appears in the Gibson table. In the C3D and manipulator tables, we can't tell whether the gains come from the learned value function or from the MPC policy. The authors had the right idea in Gibson; they should report it everywhere.\n\nThird, the baselines. NTF and P-NTF are upgraded with the same attention encoder as 'ours', which is reasonable for generalization testing, but it means the comparison isn't against the published versions. And the code isn't out yet, despite the abstract's repo link.\n\nNone of these are fatal to the central idea. The TD-Eikonal combination is sensible, the metric architecture is a legitimate contribution, and the authors' own conclusion is honest about generalization limits. The paper is useful to people working on self-supervised planning or learned cost functions. I'd give it a serious referee: ask for a corrected Eq. 6, Ours-G results everywhere, and code release.\n\nMy verdict: engage with the paper, but don't take its loss function at face value.","headline":"The paper's obstacle-normal loss (Eq. 6) is undefined where it acts due to clipping, yet the rest of the work is solid enough to warrant a careful referee rather than a desk reject.","tokens_in":19145,"tokens_out":3059,"would_cite":true,"duration_ms":31883,"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":"This paper claims that adding a temporal difference loss and a metric-space parameterization to neural Eikonal solvers yields self-supervised motion planners that succeed in complex and unseen environments, reporting up to 99.4 percent…","keywords":["robot motion planning","Eikonal equation","temporal difference learning","metric learning","self-supervised learning","neural fields","geodesic distance","Bellman optimality"],"falsifier":"Run the method on a maze with a deliberately narrow corridor, compute $d_{\\mathrm{obs}}$ with a signed-distance routine that is known to be non-smooth along the medial axis, and compare the learned travel-time contours against a dense Fast Marching reference; if the contour error concentrates exactly where the finite-difference gradient of $S^\\star$ disagrees with the analytic gradient used in the obstacle-normal loss, the differentiability assumption is the weak point. Alternatively, toggle the obstacle-normal loss in that corridor: failure to converge would indicate the term is receiving wrong normal directions.","tokens_in":18046,"feed_emoji":"🤖","tokens_out":9616,"duration_ms":89571,"temperature":0.7,"pith_summary":"This paper argues that self-supervised neural motion planners trained by solving the Eikonal equation fail in complex and unseen environments because they enforce only the equation's pointwise gradient condition and not the two further identities that a true solution carries: being an optimal value function and being a geodesic distance. The proposed remedy adds a temporal difference loss that enforces Bellman's principle over a finite time step, an obstacle-normal alignment loss, a causality weighting, and a metric-learning architecture that stores the travel time as a learned latent-space distance with an $\\ell^1$/ $\\ell^\\infty$ combination. With these modifications the paper reports a 99.4 percent success rate in cluttered 3D scenes and 91 percent for a 12-DOF dual-arm robot in a confined cabinet, with planning times often below a tenth of a second. For a general reader the upshot is that a collision-free path can be drawn from a learned cost field by sampling, without expert demonstrations and without gradient descent at runtime.","feed_headline":"Eikonal + temporal difference planner hits 99.4% success","feed_subtitle":"Self-supervised network enforces Bellman optimality and a geodesic metric to plan faster in unseen scenes, 2-12 DOF.","key_machinery":"The central object is the learned travel-time field $T(q_s,q_g)$ whose gradient norm is constrained to equal the reciprocal ground-truth speed $1/S^\\star(q)$, making it simultaneously an optimal value function and a geodesic distance. The mechanism is the composite loss $\\mathcal{L}=(\\lambda_E\\mathcal{L}_E+\\lambda_{TD}\\mathcal{L}_{TD}+\\lambda_N\\mathcal{L}_N)\\mathcal{L}_C$, where $\\mathcal{L}_E$ is the Eikonal speed loss, $\\mathcal{L}_{TD}$ enforces $T(q_s,q_g)=\\Delta t/S^\\star(q_g)+T(q_s,q_g+u_g^\\star\\Delta t)$ along the optimal direction, $\\mathcal{L}_N$ aligns the field's gradient with obstacle normals near obstacles, and $\\mathcal{L}_C=\\exp(-\\lambda_C T(q_s,q_g))$ enforces causality by prioritizing small values first. The metric space is parameterized as $T(q_s,q_g)=D(f_\\theta(q_s),f_\\theta(q_g))$ with $D(x,y)=\\sum_i \\max_j |x_{i,j}-y_{i,j}|$, an $\\ell^1$-sum of $\\ell^\\infty$ terms chosen so multiple shortest paths are not collapsed into one chord.","core_discovery":"The paper's central claim is that the Eikonal equation's solution should be read simultaneously as the optimal value function of an optimal control problem and as the geodesic distance of a Riemannian manifold whose metric is $1/S^\\star(q)$, and that a neural travel-time field trained only with the pointwise Eikonal loss can satisfy the PDE at sampled points yet drift arbitrarily between them. The paper asserts that adding a temporal difference loss derived from a Taylor expansion along the optimal policy fixes the inter-sample drift, an obstacle-normal alignment term anchors early training near obstacles, a causality weight enforces one-way value propagation, and a metric-space parameterization $T(q_s,q_g)=D(f_\\theta(q_s),f_\\theta(q_g))$ with $D$ a sum of $\\ell^\\infty$ terms preserves triangle inequality, symmetry, and non-negativity while allowing multiple shortest paths. With these ingredients and environment conditioning through attention on the obstacle point cloud, the paper reports that the learned field solves the Eikonal equation accurately enough to support sampling-based MPC inference and to generalize to unseen environments across 2 to 12 degrees of freedom.","pith_inferences":["If the differentiability assumption on $d_{\\mathrm{obs}}$ breaks at medial-axis kinks, the obstacle-normal loss in Eq. 6 should show exactly where: train the same architecture in a corridor whose width approaches the TD step and the finite-difference surrogate of $S^\\star$; expect the contour error to spike where the two gradients disagree.","The paper's metric argument suggests the $\\ell^1$/ $\\ell^\\infty$ combination is load-bearing, so a natural test is to replace only the distance $D$ with, say, a learned quasimetric while keeping every other loss term; if success falls on multi-connected workspaces, the multipath-preserving property is confirmed.","Because the paper reports weaker generalization to unseen Gibson homes, its own numbers imply the bottleneck is the environment encoder rather than the Eikonal loss; swapping the point-cloud encoder for another shape-conditioning architecture and re-measuring unseen-scene success would isolate that claim.","The TD step $\\Delta t$ is tied to environment clutter (0.02 in open scenes, 0.005 for manipulators), which predicts a direct trade-off: any cluttered environment whose narrowest passage gives a time step below what the model can represent should force either more training epochs or a smaller $\\Delta t$."],"forward_implications":["A travel-time field that is both a metric and an Eikonal solution lets the planner extract paths with sampling-based MPC alone, so inference needs no gradient computation and can recover from local inaccuracies through stochastic exploration.","Combining the Eikonal loss with the finite-step temporal difference loss is what suppresses spurious local minima: in the 2D maze ablation the full loss reaches error 0.08 versus 1.13 with only the Eikonal loss and 0.21 without the TD loss.","Environment conditioning through point-cloud attention makes the learned Eikonal solver generalizable: on unseen C3D scenes success is 99.2 percent and on unseen 7-DOF manipulator scenes 84.0 percent, close to seen-scene performance.","The same learned cost-to-go can serve other downstream planners such as cost-aware T-RRT-style sampling, because the field is a valid metric and not merely a local gradient.","In the 12-DOF real-world cabinet task the method reports 91 percent success in about 0.09 seconds on average, where prior self-supervised planners did not converge."],"supporting_citations":[{"why":"Supplies the base physics-informed neural motion planner: travel time parameterized as the Euclidean distance divided by a learned distortion, ground-truth speed from truncated obstacle distance, and the Eikonal loss $\\mathcal{L}_E$.","marker":"Ni & Qureshi, 2023a"},{"why":"Provides the temporal difference value-learning principle from which the finite-step Bellman loss $\\mathcal{L}_{TD}$ is derived.","marker":"Sutton, 2018"},{"why":"Establishes the Eikonal equation and the fast marching method used as the ground-truth travel-time solver in maze ablations.","marker":"Sethian, 1996"},{"why":"Supports the claim that Euclidean latent embeddings collapse multiple shortest paths, motivating the $\\ell^1$-of-$\\ell^\\infty$ metric distance.","marker":"Pitis et al., 2020"},{"why":"Supplies the quasimetric learning view and metric-property constraints that the network architecture enforces.","marker":"Wang et al., 2023"},{"why":"Provides the PointNext point-cloud encoder used to condition the travel-time field on the obstacle environment.","marker":"Qian et al., 2022"},{"why":"Supplies the causality weighting scheme that prioritizes smaller travel times during training.","marker":"Wang et al., 2024b"},{"why":"Underwrites the differentiability assumption for $d_{\\mathrm{obs}}$ through differentiable forward kinematics.","marker":"Villegas et al., 2018"},{"why":"Provides the sampling-based MPC used at runtime for path inference.","marker":"Williams et al., 2016; Bharadhwaj et al., 2020"}],"fun_headline_variants":["TD metric Eikonal planner hits 99.4% success","Temporal difference metric learning sharpens Eikonal planning","Eikonal solver with TD metric learning generalizes to unseen scenes","New loss fixes Eikonal drift, enabling complex motion planning","Bellman optimality plus geodesic metric for better robot path"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training pipeline assumes that the minimum workspace distance between the robot's geometry at configuration $q$ and the obstacles, $d_{\\mathrm{obs}}(q,\\mathcal{X}_{\\mathrm{obs}})$, is differentiable in $q$ through differentiable forward kinematics; if that distance is kinked or approximated in narrow passages or high-DOF arms, the Eikonal, temporal difference, and obstacle-normal losses carry undefined or biased gradients.","fun_headline_variants_meta":{"raw":{"variants":["TD metric Eikonal planner hits 99.4% success","Temporal difference metric learning sharpens Eikonal planning","Eikonal solver with TD metric learning generalizes to unseen scenes","New loss fixes Eikonal drift, enabling complex motion planning","Bellman optimality plus geodesic metric for better robot path"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001062,"raw_usage":{"total_tokens":4461,"prompt_tokens":958,"completion_tokens":3503,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":3416}},"tokens_in":574,"tokens_out":3503,"duration_ms":24547,"temperature":1.0,"reasoning_tokens":3416,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:00:11.365148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a maze with a deliberately narrow corridor, compute $d_{\\mathrm{obs}}$ with a signed-distance routine that is known to be non-smooth along the medial axis, and compare the learned travel-time contours against a dense Fast Marching reference; if the contour error concentrates exactly where the finite-difference gradient of $S^\\star$ disagrees with the analytic gradient used in the obstacle-normal loss, the differentiability assumption is the weak point. Alternatively, toggle the obstacle-normal loss in that corridor: failure to converge would indicate the term is receiving wrong normal directions.","supporting_citations":[{"cited_title":"A fast marching level set method for monotonically advancing fronts","cited_arxiv_id":null,"evidence_quote":"Establishes the Eikonal equation and the fast marching method used as the ground-truth travel-time solver in maze ablations."},{"cited_title":"Neural kinematic networks for unsupervised motion retargetting","cited_arxiv_id":null,"evidence_quote":"Underwrites the differentiability assumption for $d_{\\mathrm{obs}}$ through differentiable forward kinematics."},{"cited_title":"Aggressive driving with model predictive path integral control","cited_arxiv_id":null,"evidence_quote":"Provides the sampling-based MPC used at runtime for path inference."}],"review_version":1}