acceptodds
Under review as a conference paper at ICLR 2027

Learning Neural Time Fields on Randers Manifolds for Drift-Aware Motion Planning

Abstract

Physics-informed neural time fields provide a self-supervised learning framework for robot motion planning. Existing Eikonal and Riemannian formulations primarily model isotropic or symmetric motion costs. This limits their ability to represent directional effects arising from environmental flows, gravity, and direction dependent dynamics. We introduce Randers Neural Time Fields (RaNTFields), a physics-informed neural framework for learning asymmetric anisotropic travel-time fields. RaNTFields extends the Eikonal formulation using a Randers metric that combines a configuration-dependent Riemannian metric with a directional one-form, derived from the corresponding Hamilton-Jacobi equation. We evaluate RaNTFields on 2D navigation under complex flows, terrain navigation, and 5-DoF robot arm motion generation under gravitational drift. In constant-flow environments, RaNTFields achieves a 0.997 correlation with the Ordered Upwind reference solution across start-goal pairs. In the 5-DoF arm, it follows the preferred gravity drift direction and consequently reduces the peak torque by relative to an Eikonal planner, with a success rate compared to for the baseline. These results demonstrate that RaNTFields can encode directional preferences arising from environmental flows and robot dynamics.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot motion planning, Randers Equation, Physics-informed Neural Networks, Differential Geometry

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_m45AOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RaNTFields, a physics-informed neural framework for asymmetric motion planning based on Randers geometry. The method learns drift-aware travel-time fields through self-supervision, without requiring trajectory supervision. The fields can then be used to compute paths that navigate the robot to the goal, in the presence of obstacles and directional environmental factors (e.g., gravity, flows). The experiments demonstrate strong agreement with numerical Ordered Upwind Method (OUM) solutions and reduced peak torque in robotic arm planning.

Critical strengths
  • The paper tackles an interesting problem in robot motion planning, dealing with the case where the motion plan depends on asymmetric environmental factors, including flows, gravity, and directional dynamics.
  • The proposed formulation is relatively simple and elegant: add a drift term that introduces asymmetries and learn a physics-informed network that can learn the time-travel fields in a self-supervised manner.
  • The presentation is very clear and provides required preliminaries before introducing the contribution.
  • The experiments are convincing and include results on a real robot.
Critical weaknesses
  • It is unclear if the result in Section 4.1 (in particular, eq (5)) is a novel/non-trivial result. I understand this is simply obtained by replacing the Mahalanobis norm in (3) with a Randers metric, but the first sentence of Section 4.1 hints to the fact that there is more to the derivation. If so, I suggest formalizing the result in a theorem/proposition.
  • I understand that G, S*, f are given and are not learned. In practice, how should one set G and f (I guess S* is the same as eq (2)).
  • In practice, would one need to retrain the network every time an obstacle changes?
  • The UR5e experiments compare RaNTFields against Neural Eikonal, Neural Riemannian, RRT, and RRT*, but none of these baselines explicitly optimizes actuator torque. Since reducing peak torque is a central experimental claim, comparisons against torque-aware trajectory optimization methods would strengthen the evaluation. One potential baseline is Crocoddyl (Mastalli et al., Crocoddyl: An Efficient and Versatile Framework for Multi-Contact Optimal Control, ICRA 2020), which supports optimal control formulations with actuation-related costs. Collision avoidance could be incorporated using Colmpc (https://github.com/agimus-project/colmpc).
  • Missing architectural ablation studies: The proposed architecture combines Interval Quasimetric Embedding (IQE) with a residual MLP, but the contributions of these components are not evaluated separately. Since a standard MLP can already represent asymmetric travel-time functions, it is unclear whether IQE provides substantial benefits. I suggest comparing against a vanilla MLP trained with the same Randers PDE loss, as well as IQE-only and residual-only variants, to isolate the contributions of each architectural component.
  • The current experiments are relatively simple, consisting primarily of synthetic 2D flow fields and a 5-DoF robotic arm with a single spherical obstacle. It remains unclear how well RaNTFields generalizes to more challenging environments (including multiple obstacles and narrow passages) and higher-dimensional systems. Evaluating such scenarios would better demonstrate the method's robustness and practical applicability.

Minor weaknesses:

  • Several grammatical errors should be corrected: Page 1, Introduction: "provide strong theoretical guaranties" should be "provide strong theoretical guarantees.". Page 7, Section 5.1.1: "These results are consistent among the three flow environments and they demonstrate that…" should be "These results are consistent across the three flow environments and demonstrate that…". Page 6, Section 4.3: "such that we recover the identitity riemannian metric with there is no elevation" should be "such that the identity Riemannian metric is recovered when there is no elevation."
  • Unclear definition of active torque. The definition of "Active torque" in Table 3 is not sufficiently clear. The authors should explicitly explain how this metric is computed and how it differs from mean and peak torque.
  • Insufficient implementation details for baselines: The paper does not clearly specify which codebases or implementations were used for the baseline methods, or whether these baselines were reimplemented by the authors. Providing implementation details, source-code references, and relevant hyperparameters would improve reproducibility and help ensure fair comparisons.
Questions for authors
  • how is eq (5) obtained? why is it a nontrivial replacement of the Mahalanobis norm in eq. (3)?
  • I understand that G, S*, f are given and are not learned. In practice, how should one set G and f?
  • Would one need to retrain the network every time an obstacle changes?
  • can you add a comparison against torque-aware baselines for the UR5e experiment?
  • did you perform any ablation experiments on the proposed architecture?
  • how would the 5-dof robot arm experiment generalize to a more complex environment? how does RaNTfields scale to higher-dimensional systems?
  • what is the definition of active torque in Table 3?
1 comment

We thank the reviewer; see also the general response.

Q1. How Eq. (5) is obtained, and why it is more than a norm swap

It is the stationary HJB equation of the Zermelo problem q̇ = f(q) + u, uᵀG(q)u ≤ 1, with running cost 1/S(q):

0 = min over uᵀGu ≤ 1 of [1/S + ∇Tᵀ(f + u)].

The minimum of a linear function over the G-unit ball is −√(∇TᵀG⁻¹∇T), which gives Eq. (5): √(∇TᵀG⁻¹∇T) − fᵀ∇T = 1/S (see App. A.1). Three things make it more than a norm replacement:

  • The drift enters the Hamiltonian linearly, as a one-form, not inside the norm. Writing the same problem as a path cost gives the Randers metric F(y) = √(yᵀa y) + bᵀy with a = (λG + GffᵀG)/λ², b = −Gf/λ, λ = 1 − fᵀGf. A norm swap is correct only with this a and b.
  • Well-posedness needs |f|_G < 1. Otherwise λ ≤ 0, a is not positive definite, and some directions become unreachable (T = ∞). This is the feasibility condition we enforce when constructing f.
  • Eikonal (G = I, f = 0) and Riemannian (f = 0) are recovered as special cases.

This is now Proposition 1 in Sec. 4.1, with the proof in App. A.1-A.2.

Q2. How to set G and f

G sets how costly motion is in each direction; f sets the preferred direction. In our experiments:

  • Flows: G = I and f is the current's velocity field.
  • Terrain with height h: G = I + ∇h∇hᵀ (the surface metric) and f = −c∇h/(‖∇h‖ + ε), downhill.
  • Arm: G = M(q)/m_ref + λI (inertia metric) and f ≈ −τ_g/m_ref (up to scale) in normalized joint coordinates, derived from gravity.
  • User-specified: a workspace direction w (e.g. "up") pulled back to joint space, either f ∝ J(q)ᵀw or gravity with its direction replaced by w (general response, Sec. 2).

In every case the feasibility condition fᵀGf < 1 holds everywhere: flow speeds are chosen below 1, and terrain and arm drifts are scaled once (max 0.9 and 0.95).

Q3. Retraining when obstacles change

Yes. Like NTFields and TD-NTFields, a field is trained per environment. Our change is the PDE, so environment generalization is orthogonal: conditioning the field on an environment encoder (e.g. a PointNet++ embedding of the obstacle point cloud) with a larger backbone is a natural route. This is future work, and the paper now states it as a limitation.

Q4. Torque-aware baseline (Crocoddyl + Colmpc)

We ran Crocoddyl with Colmpc constraints on all 23 arm spheres, on the same 100 pairs, with RaNTFields' dynamics model and time to goal, tracked in the same MuJoCo harness.

The two methods solve different problems. Crocoddyl is dynamics-aware (over q, q̇ and τ) and minimizes torque directly. RaNTFields is a kinematic, first-order travel-time field over q: inertia only shapes G and gravity only sets f. We also solved RaNTFields' own objective exactly (kinematic OCP, same constraints) and executed it with our controller.

On the same pairs (revised paper, Table 10): RaNTFields 92% success, 75.0 N·m peak torque, ~4.3 ms per query; Crocoddyl + Colmpc 100%, 47.3 N·m, 6.9 s median per solve; kinematic OCP 98%, 75.1 N·m, 35-79 s per solve.

The exact optimum of our kinematic objective gets the same torque as RaNTFields, so the gap to Crocoddyl is not due to the quality of the learned field, but to the different problem it solves. RaNTFields is a real-time field encoding a preferred direction, not a torque optimizer; its torque reduction over the Eikonal field comes from G, not f (general response, Sec. 3). We revised the torque claims accordingly (Sec. 5.2, App. A.6.5).

Q5. Ablation

Yes: IQE-only, residual-only and a vanilla MLP on the same Randers loss, capacity-matched, plus the TD-NTFields metric network (general response, Sec. 1). Both branches are needed; IQE gives a modest gain over a symmetric distance.

Q6. Harder environments and higher dimensions

We added a 2D two-corridor scene with opposing currents (general response, Sec. 2). We also tried a 6-DoF UR5e in a cluttered cabinet-shelf scene. Without the drift, our pipeline reaches 74% success (TD-NTFields' Eikonal recipe: 82%). With the gravity drift on the same backbone, success drops to ~51%, so learning the drift, not clutter, is a bottleneck. Possible directions for scaling to higher dimensions and clutter are giving the network f(q) as an input feature, an environment encoder (as in Q3), and a more expressive architecture. The paper now states the difficulty as a limitation and adds these directions as future work.

Q7. Active torque

Active torque = √(∫(τᵀq̇)² dt), the root-integrated-square of mechanical power (torque that does work). All torque metrics are now defined in App. A.5.

Minor points

We fixed the three typos. Baselines: N-Eik. and N-Riem. use our exact pipeline (data, network, controller and gains); N-Eik. sets f = 0 and G = I, N-Riem. sets f = 0. RRT and RRT* are OMPL's implementations (2 s planning budget, the same collision checker). These details are now in App. A.5; the code will be released with the camera-ready version.

Reviewer_nA7DOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RaNTFields, a physics-informed neural time-field framework for motion planning with asymmetric, direction-dependent costs. Building on neural Eikonal and Riemannian formulations, the authors introduce a Randers Hamilton–Jacobi equation that incorporates a configuration-dependent metric and a drift field. The network combines an IQE-based quasimetric branch with a residual MLP and is trained using PDE constraints without ground-truth travel-time labels.

The method is evaluated on 2D flow and terrain navigation tasks against Ordered Upwind Method (OUM) solutions, and on a 5-DoF UR5e manipulator. The 2D experiments show good agreement with numerical references in terms of gradient directions and travel-time asymmetry. The manipulator experiments report reduced peak torque compared with Eikonal and Riemannian baselines, along with a qualitative hardware demonstration.

Critical strengths
  1. A meaningful extension of neural time fields to asymmetric motion costs.

The paper addresses a limitation of existing Eikonal and Riemannian neural planners: their inability to represent different costs for forward and reverse motions. Introducing a drift term into the travel-time formulation is well motivated and provides a principled way to model directional preferences.

  1. Encouraging agreement with numerical reference solutions.

The 2D experiments compare the learned fields against OUM solutions across several flow and terrain settings. The reported gradient-direction similarities and asymmetry correlations suggest that the network can approximate the intended asymmetric field without requiring precomputed optimal travel-time labels. This is a useful result, even if it does not by itself establish the optimality of the extracted paths.

Critical weaknesses

My main concern is the limited differentiation from NTFields, TD-NTFields, and recent Riemannian neural distance-field approaches.

The overall framework remains largely within the established physics-informed neural time-field paradigm. The causality weighting is explicitly adapted from TD-NTFields, the Randers formulation follows classical geometric optimal-control results, and the asymmetric representation uses an existing IQE construction. While combining these elements is reasonable, it is unclear which part of the proposed architecture provides a substantive advantage beyond replacing the symmetric PDE with its drift-aware counterpart.

In particular, the experiments do not isolate the contributions of the IQE branch, residual branch, and their learned combination. Nor do they compare against a simpler asymmetric neural field trained with the same Randers PDE and comparable training settings. Such a comparison is important because the asymmetric PDE itself may account for most of the observed improvements.

I recognize that the original symmetric TD-NTFields formulation cannot directly represent asymmetric travel times. My concern is therefore not simply the absence of an unchanged TD-NTFields baseline, but the lack of evidence separating the benefit of the new physical formulation from that of the proposed network architecture and training design.

As a result, I find the methodological novelty and necessity of the full proposed framework insufficiently demonstrated.

Questions for authors

Q1. What is the essential methodological contribution beyond adapting previous neural time-field formulations to the Randers-HJB equation?

Could the authors clarify the differences from TD-NTFields and the recent Riemannian neural distance-field approach at the level of architecture, training objective, and learned representation?

More importantly, how does the complete RaNTFields architecture compare against a simpler asymmetric MLP or an appropriately adapted NTFields/TD-NTFields-style model trained with the same Randers PDE? An ablation of the IQE and residual branches would help establish whether the proposed architecture is necessary for the reported gains.

Q2. Can the authors provide quantitative real-robot validation of the claimed torque benefits?

Figure 4 shows a hardware execution example, but Table 3 appears to report simulation results. Are any repeated real-robot trials available comparing RaNTFields with the Eikonal and, especially, the Riemannian baseline on matched start-goal tasks?

Measurements of peak and mean joint torque, execution success, and trajectory duration would be particularly informative. It would also be helpful to clarify whether the reported simulation torque statistics are computed over comparable trajectories and whether the same controller parameters are used across the neural baselines.

This question matters because the additional peak-torque reduction over the Riemannian model is relatively modest, and the claimed practical benefit is a major motivation for introducing drift-aware planning.

1 comment

We thank the reviewer. The main concern, separating the PDE from the architecture, is answered by the new ablation (general response, Sec. 1).

Main concern and Q1: PDE vs architecture

We agree the components exist: the Randers metric is classical, IQE is from quasimetric RL, and causality weighting is from TD-NTFields. We separate their effects on two axes:

PDE alone. Even solved exactly on a grid (no network), the symmetric PDEs cannot represent T(s→g) ≠ T(g→s): their travel time is 34-61% off OUM-R, vs 2-13% for our network (general response, Sec. 2). This error belongs to the PDE, so no architecture trained on it can remove it. The Randers PDE is what makes the asymmetry learnable at all.

Architecture, PDE fixed. Same Randers PDE, recipe and capacity; 2D: 5 scenes × 3 seeds; arm: 3 seeds.

Head (same Randers PDE)2D success ↑2D TT err ↓Arm success ↑
Full (IQE + residual) (ours)0.97 ± 0.060.0890 ± 1%
Simple asymmetric MLP (plain PINN)0.74 ± 0.240.1213 ± 6%
Residual only0.79 ± 0.180.1220 ± 5%
IQE only0.85 ± 0.160.2578 ± 1%
TD-NTFields head, made asymmetric0.87 ± 0.100.1280 ± 0%
TD-NTFields head, as published0.50 ± 0.230.4855 ± 8%

TT err: median |T − T*|/T* vs OUM-R.

  • With the PDE fixed, the architecture adds 10-23 points of 2D success and 10-77 points on the arm.
  • The simple asymmetric MLP and the adapted TD-NTFields head both solve the same PDE, but worse. The published TD-NTFields head is symmetric, so it cannot represent the asymmetry at all.
  • IQE alone is unstable (4/15 runs collapsed) and the residual alone stalls on the arm; the gated combination (α ≈ 0.49) fixes both.

So the PDE enables the asymmetry and the architecture makes it learnable reliably; neither alone gives the results.

Differences at the three levels requested:

NTFieldsTD-NTFieldsRiemannian NESOurs
ArchitectureSymmetric operator, T = ‖q_s − q_g‖/τPirateNet + L1-of-L∞ metric embeddingSymmetrized, factorized TGated IQE quasimetric + residual MLP
ObjectiveEikonalEikonal + TD + normal alignment + causalityRiemannian eikonalRanders HJB + causality
RepresentationSymmetricSymmetric metricSymmetricAsymmetric

Here the Randers HJB is √(∇Tᵀ G⁻¹ ∇T) − fᵀ∇T = 1/S, well posed when |f|_G < 1, with Eikonal (G = I, f = 0) and Riemannian (f = 0) as special cases.

Essential contribution: (a) planning with a preferred direction of motion as Zermelo navigation and its Randers HJB; (b) an asymmetric solver, with evidence that symmetric heads cannot solve it and asymmetric ones need both branches; (c) constructing G and f for currents, slopes, gravity and a user-chosen direction (general response, Sec. 4).

Q2: Torque evidence

Hardware. We ran the upward-drift field and N-Riem. closed-loop on a real UR5e (field queried at the encoder reading, velocity commands at 222 Hz). On 8 start-goal pairs selected for safety only, all 16 runs reached the goal, and the end effector was a median 0.045 m higher with the upward drift (6/8 pairs, p = 0.039), matching simulation (0.041 m, per-pair r = 0.98). Torque numbers stay from MuJoCo; we label Fig. 4 qualitative and restrict torque claims to simulation.

Comparable trajectories. All three fields are run by the same controller with identical gains (cruise speed, λ, goal tolerance); the drift-free baselines simply have f = 0. Same 100 start-goal pairs, statistics over the 91 pairs all three complete. With the corrected drift (general response, Sec. 3):

ModelSuccessTime to goal (s)Peak torque (N·m)Mean torque (N·m)
RaNTF92%11.35 ± 3.1375.0 ± 29.17.45 ± 1.38
N-Riem.99%10.62 ± 2.3477.9 ± 33.57.41 ± 1.41
N-Eik.98%10.39 ± 2.3490.2 ± 34.87.65 ± 1.37

RaNTF keeps the lowest peak torque and matches N-Riem.'s mean torque. Its longer time to goal is expected: the field follows the preferred direction rather than the shortest route, as in the 2D flows.

We agree the gain over Riemannian is modest. Paired Wilcoxon tests, RaNTF − N-Riem.: peak −1.7 N·m (p = 0.12), mean +0.04 N·m (p = 0.62), both not significant; time +0.4 s (p = 8e-11). The peak-torque gain over Eikonal (−12.8 N·m, p = 1e-7) is shared by N-Riem., so it comes from the metric G, not the drift f.

We corrected the attribution in the paper (abstract and Sec. 5.2): the drift encodes a preferred direction, not a torque reduction. On the arm, motion is more aligned with gravity than N-Riem. with the same G (p = 4e-7), and a user-chosen "up" direction lifts the end effector by 0.042 m (p = 1e-11). The main evidence for the drift is the 2D results, where it changes both the travel time and the route. All of these results are in general response, Sec. 2.

Reviewer_eYMwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper extends physics-informed neural time fields (NTFields and follow-ups) from Eikonal and Riemannian costs to Randers costs, so that the travel time from A to B can differ from B to A. The motivating cases are navigation in a flow, terrain where uphill and downhill cost differently, and an arm moving with or against gravity. The problem is posed as Zermelo-type navigation with a configuration-dependent metric and a drift field, and the corresponding stationary Hamilton-Jacobi equation is derived. A goal-conditioned network is trained to satisfy this equation at sampled configuration pairs, so no travel-time labels are needed. The network mixes an interval quasimetric embedding (IQE) branch, asymmetric by construction, with a residual MLP branch through a learned gate, and its gradient gives a feedback direction at test time. Experiments cover 2D constant, circular and saddle flows and two terrains, compared with Ordered Upwind (OUM) solutions under Randers, Riemannian and Eikonal costs, and a 5-DoF UR5e arm with a gravity-derived drift, compared with neural Eikonal and Riemannian versions of the same pipeline and with RRT/RRT*. In 2D the learned field closely follows the Randers OUM solution in gradient direction and in the forward/reverse asymmetry. On the arm, the Randers planner lowers peak torque relative to the Eikonal and Riemannian versions, with slightly longer paths and a slightly lower success rate.

Critical strengths
  1. The Eikonal, Riemannian, Randers hierarchy is clear, and the derivation from the optimal control problem to the Randers form (App. A.1-A.2) is correct and easy to follow. The drift feasibility condition is tied properly to positivity of the metric, and setting the drift to zero (and the metric to identity) recovers the earlier NTFields formulations as special cases.
  2. Training stays label-free. Only local evaluations of speed, metric and drift are needed, so the practical appeal of NTFields carries over to direction-dependent costs.
  3. The paper is well structured and easy to follow.
Critical weaknesses
  1. The reviewer believes the novelty is incremental. The PDE is the classical Zermelo/Randers construction, training follows NTFields/TD-NTFields, and the arm controller's goal pull is from Li et al. (2026). The new piece is the IQE plus residual network, and it is not ablated, although TD-NTFields (Ni et al., 2025) already tested IQE as a latent metric and found it worse than its own (0.32 vs 0.08 maze error). The learned gate is never reported, and the residual branch and pair-dependent scale drop the zero-diagonal and triangle-inequality properties used to motivate IQE. Eik-HiQRL (Giammarino and Qureshi, ICLR 2026), which already pairs a quasimetric value function with an Eikonal residual, is not discussed.
  2. The 2D study does not measure solver quality. OUM-G and OUM-E solve symmetric problems, so their zero asymmetry holds by construction and their lower cosine is expected. No other solver of the same Randers PDE is compared (e.g. a plain PINN or the TD-NTFields network with the same residual). "Length" is Euclidean although the objective is travel time, and the travel cost of extracted paths versus the OUM optimum is not reported.

Minor comments:

  1. There are sections 2.0.1, 2.1, 2.2. It should be 2.1, 2.2, 2.3
  2. The number of citations is too few.

• Ni, R., Pan, Z., and Qureshi, A. H. Physics-Informed Temporal Difference Metric Learning for Robot Motion Planning. ICLR 2025. arXiv:2505.05691. • Giammarino, V. and Qureshi, A. H. Goal Reaching with Eikonal-Constrained Hierarchical Quasimetric Reinforcement Learning. ICLR 2026. arXiv:2512.12046. • Mirebeau, J.-M. and Portegies, J. Hamiltonian Fast Marching: A Numerical Solver for Anisotropic and Non-Holonomic Eikonal PDEs. Image Processing On Line, 9:47-93, 2019. • Jaillet, L., Cortés, J., and Siméon, T. Sampling-Based Path Planning on Configuration-Space Costmaps. IEEE Transactions on Robotics, 26(4):635-646, 2010.

Questions for authors
  1. What is the contribution beyond combining the classical Randers Hamilton-Jacobi equation with NTFields-style training? The only new component is the IQE plus residual network, so please report the learned gate after training and ablate the network on the 2D tasks and the arm: IQE only, residual only, and the TD-NTFields metric without its symmetric operator, all trained on the same Randers residual, with travel-time error against OUM-R and asymmetry correlation. TD-NTFields found IQE worse than its own metric on symmetric mazes, so it is important to show that IQE actually helps for asymmetric fields and that the residual MLP is not doing most of the work.
  2. The residual branch is nonzero at q = q_g, and the pair-dependent scale can break the triangle inequality. Does the final field satisfy T(q, q) = 0 and the triangle inequality in practice? Please report the boundary error and the fraction of sampled triples that violate the triangle inequality. If these properties do not hold, what does IQE add beyond a generic asymmetric parameterization?
  3. For the 2D tasks, please report (a) the travel time of extracted paths under the true Randers metric, normalized by the OUM-R optimum, since this is the quantity the method optimizes, and (b) at least one other solver of the same Randers PDE, e.g. a plain PINN or the TD-NTFields network trained with your residual. Without (b), Tables 1 and 2 show that the Randers problem differs from the symmetric ones, not that RaNTFields solves it better than the alternatives.
2 comments

We thank the reviewer. The ablation, gate and solver results are in the general response; we summarize them and add new checks.

Q1. Contribution, gate, ablation

Contribution. We agree the Zermelo/Randers HJB is classical, and we do not claim the goal-pull controller (Li et al., 2026). Our contribution is (a) posing planning with a preferred direction of motion as Zermelo navigation with metric G and drift f, solved label-free; (b) showing that symmetric neural time fields cannot represent it and that a two-branch asymmetric network can (below); (c) building f from currents, slopes, gravity, or a user-chosen workspace direction (general response, Sec. 2).

Gate. α = 0.48-0.51 in every scene, seed and variant (initialized at 0.5). The paper now reports it (Sec. 4.2.2) and describes the blend as near-even rather than learned selection.

Ablation. The table in general response Sec. 1 has every variant you asked for, trained on the same Randers residual, with travel-time error and asymmetry correlation against OUM-R. That includes the TD-NTFields metric with its symmetric operator replaced by relu(e_s − e_g) (success 0.87 vs our 0.97; 0.50 as published). The residual MLP is not doing most of the work: alone it is among the weakest variants (success 0.79), and neither branch is sufficient.

On TD-NTFields' IQE result: we see the same. IQE alone collapsed on 4/15 runs, and IQE on TD-NTFields' embedding was worse than its relu distance (success 0.53 vs 0.87). IQE helps only combined with the residual branch. Replacing IQE by a symmetric L1 distance inside the full model is close on average; IQE gives the better gradient field in 15/20 paired runs (sign test p = 0.04), most clearly under strong uniform flow. The paper states it as a modest gain.

Q2. T(q, q) = 0 and the triangle inequality

4000 free-space points and triples per scene; means over all 15 runs:

|T(q,q)| / median TRandom triples violated (>5%)Near-collinear triples violated (>5%)Median excess
Full0.09 ± 0.04%1.3% (0.14%)11% (3.1%)2.5%
Plain MLP0.42 ± 0.12%2.6% (0.64%)20% (9.5%)4.8%
TD-NTFields, relu0.13 ± 0.04%1.9% (0.34%)19% (9.1%)4.7%
Pure IQE output0000

Near-collinear: b near the a-c midpoint, where a true travel time is close to equality.

The reviewer is right that the full field is not a quasimetric by construction. In practice the boundary error is ~0.1% of a typical travel time, and triangle violations are rare and small: the median excess (2.5%) is below the field's own travel-time error (7.7%). It violates about 2× less often than the plain MLP.

What IQE adds: an asymmetric distance with an exact zero diagonal, which keeps the field near-quasimetric and gives a modest accuracy gain. A pure IQE output guarantees both properties but is much less accurate (success 0.53). The revised Sec. 4.2.2 presents IQE as an inductive bias, not a guarantee, and App. A.6.4 reports these checks.

Q3. (a) Randers travel time of extracted paths

Each path is scored under the true Randers metric used by OUM-R and divided by the cost of OUM-R's own path (same integrator, so grid error cancels). The median ratios per scene are the Cost columns of Table 2 in the revised paper; the two-lane scene is in Table 7.

Our paths are within 0.4-2.9% of OUM-R's. Symmetric planners pay a large price only where the drift changes the optimal route: 23% in circular flow and 2.15× in the new two-lane scene, where they must use the same corridor in both directions. Elsewhere the optimal route is nearly the same straight line, so the drift changes the travel time, not the path. That is where our predicted travel time matters: 2-8% error vs 34-61% for the symmetric planners (general response, Sec. 2). Table 2 now scores paths by Randers travel time instead of Euclidean length.

Q3. (b) Another solver of the same PDE

The plain MLP is a plain PINN, and the TD-NTFields network was trained with our residual (Q1). Both are below the full model on every pooled metric.

Related work and minor points

Eik-HiQRL constrains an IQE quasimetric value with a unit-speed, isotropic Eikonal equation for offline goal-conditioned RL, so its asymmetry comes only from the embedding. In ours, the Randers PDE itself is asymmetric, through a configuration-dependent metric and drift, and is solved without data. Hamiltonian Fast Marching is a grid solver for anisotropic eikonal PDEs, like our OUM-R reference, and T-RRT is a sampling-based planner on configuration-space costmaps. We fixed the section numbering (2.1-2.3) and added the suggested citations, with a comparison table (Table 1).

I appreciate the detailed response of authors. My concerns have been resolved and I will keep my score as is.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to all reviewers and Chairs

General response

We thank the reviewers. We ran the requested ablation and solver comparison, and narrowed our arm claims.

1. Architecture ablation (all; revised paper: Sec. 5.3, Table 4, App. A.6.4)

Same encoder, Randers loss and recipe; ~850k params each; 5 scenes × 3 seeds, 100 pairs per run, scored against the numerical Randers solution (OUM-R). Mean ± std over 15 runs.

VariantSuccess ↑cos(∇T, ∇T*) ↑Asym. err ↓Asym. corr ↑TT err ↓
Full (IQE + residual)0.97 ± 0.060.95 ± 0.060.23 ± 0.150.97 ± 0.080.08 ± 0.04
Full, L1 instead of IQE0.97 ± 0.020.96 ± 0.020.23 ± 0.110.98 ± 0.020.09 ± 0.05
Full, no distance0.81 ± 0.170.77 ± 0.130.57 ± 0.300.88 ± 0.130.12 ± 0.06
IQE only0.85 ± 0.160.80 ± 0.210.58 ± 0.420.73 ± 0.390.25 ± 0.25
Residual only0.79 ± 0.180.78 ± 0.140.54 ± 0.310.89 ± 0.120.12 ± 0.07
Plain MLP (PINN)0.74 ± 0.240.86 ± 0.160.45 ± 0.300.89 ± 0.190.12 ± 0.07
TD-NTFields head (asym.)0.87 ± 0.100.88 ± 0.080.42 ± 0.280.95 ± 0.070.12 ± 0.05

Asym.: T(s→g) − T(g→s) vs OUM-R. TT err: median |T − T*|/T*.

  • Both branches are needed: removing either costs 12-23 points of success. The residual alone is among the weakest.
  • IQE alone is unstable (collapsed in 4/15 runs), as TD-NTFields found; it works combined with the residual.
  • IQE vs L1: close on average; IQE gives a better gradient in 15/20 paired runs (sign test p = 0.04), clearest under strong uniform flow. A modest gain, not essential.
  • Gate α ≈ 0.49 in every run; the paper now calls it a near-even blend.
  • Other Randers solvers (eYMw, nA7D): the plain MLP is a plain PINN; TD-NTFields' head is symmetric as published (success 0.50) and below ours made asymmetric.

Arm (5-DoF UR5e). Same recipe and checkpoint as the paper's arm table, capacity-matched (3.8M params), 3 seeds × 100 pairs:

VariantSuccess ↑Failures per seed: stalls / collisions
Full (IQE + residual)90 ± 1%1 / 9
IQE only78 ± 1%0 / 22
Residual only20 ± 5%68 / 12
Plain MLP13 ± 6%68 / 19
TD-NTFields head (asym.)80 ± 0%7 / 13
TD-NTFields head (published)55 ± 8%6 / 39

Without the IQE branch the field mostly stalls before the goal; without the residual branch it collides more often.

2. The drift encodes a preferred direction

Travel along f is cheaper than against it, and the field learns this:

  • Flows: predicted travel time within 2-8% of OUM-R; drift-free planners are 34-61% off. In a new two-lane scene with opposing currents, symmetric planners take 2.15× the optimum (median); ours 1.07-1.09×, never the wrong corridor (0/75; Sec. 5.1.1, App. A.6.3).
  • Terrain: downhill is cheaper; error 9-13% vs 46-50% without drift.
  • Arm, gravity: motion is more aligned with gravity than the Riemannian field with the same G (paired Wilcoxon, 91 pairs, p = 4e-7).
  • Arm, user-chosen: replacing gravity in the drift by an upward workspace force lifts the end effector 0.042 m vs Riemannian (higher on 84% of pairs, p = 1e-11), and 0.045 m on a real UR5e (p = 0.039, 8 pairs; Sec. 5.2). The controller follows ∇T only, so this comes from the field.

3. Corrected arm results

  • Correction: the controller uses q̇ ∝ f(q) + u*, u* = −G⁻¹∇T/‖∇T‖_{G⁻¹}. In the published table its f was scaled differently from the trained f. With the trained f: success 96 → 92%, peak torque 74.2 → 75.0 N·m, still lowest (Riemannian 77.9, Eikonal 90.2; both drift-free, unaffected).
  • Torque comes from G, not f: the gain over Eikonal (median −12.8 N·m, p = 1e-7) is shared by Riemannian; RaNTF vs Riemannian is not significant (−1.7, p = 0.12). f instead sets the preferred direction: motion follows gravity, or a user-chosen direction (Sec. 2).
  • Torque-optimal control (Crocoddyl + Colmpc): 47 N·m, but seconds per solve vs ~4.3 ms per field query (reply to m45A, Q4).

4. Contributions

  • (a) Formulation: planning with a preferred direction as Zermelo navigation, q̇ = f + u, uᵀGu ≤ 1. G sets the direction-dependent cost, f the preferred direction. Its travel time solves a Randers HJB, well posed when |f|_G < 1, with Eikonal and Riemannian as special cases (now Proposition 1).
  • (b) Solver: a physics-informed solver that represents T(s→g) ≠ T(g→s), supported by Sec. 1.
  • (c) Specifying G and f: currents, slopes, gravity, or a user-chosen workspace direction.
MethodPDECost of motionPreferred directionTravel time
NTFields / P-NTFieldsEikonalIsotropicNoneSymmetric
TD-NTFieldsEikonal + TDIsotropicNoneSymmetric
NES / C-NESRiemannianDirection-dependent (G)NoneSymmetric
OursRanders HJBDirection-dependent (G)Drift fAsymmetric

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.