Physics-informed Neural Motion Planning via Domain Decomposition in Large Environments

Ahmed H. Qureshi; Alexiy Buynitsky; Ruiqi Ni; Yuchen Liu

REVIEW 3 major objections 5 minor 1 cited by

Physics-informed Neural Motion Planning via Domain Decomposition in Large Environments

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Latent-distance neural fields scale physics-informed motion planning to large real-world maps, planning under 0.1 seconds with high success.

desk verdict Solid empirical scaling of physics-informed neural planners, but the latent-distance representation's boundary behavior needs analysis before the method is fully trusted. read the letter →

arxiv 2506.12742 v1 pith:7NFBJKYM submitted 2025-06-15 cs.RO

Yuchen Liu , Alexiy Buynitsky , Ruiqi Ni , Ahmed H. Qureshi This is my paper

classification cs.RO

keywords motionplanningphysics-informedneuralnetworksEikonalequationdomaindecompositionfieldslatentspacedistancecost-to-gofunctionquadrupedrobot

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that physics-informed neural motion planners can scale to large environments if the cost-to-go (arrival time) is represented as a distance between latent embeddings of the start and goal, rather than as a pointwise scalar field. To do this, it partitions the environment into overlapping subdomains, learns a small neural encoder per subdomain, and computes the arrival time as the log-sum-exp of the absolute difference between the start and goal embeddings. The authors argue that this latent-space distance preserves global spatial coherence while each subnetwork stays small enough to avoid spectral bias and the difficult PDE loss landscape. If correct, a robot can plan paths in maps spanning hundreds to nearly two thousand square meters in about a tenth of a second with high success, without needing expert demonstrations, and the paper demonstrates this on a real quadruped robot.

What carries the argument

The central object is the composite latent embedding $\Phi_w(q)=\sum_i w_i\Phi_i(q)/\sum_i w_i$, where each $\Phi_i$ is a SIREN network with Fourier-feature input, normalized to its subdomain, and $w_i$ is a Gaussian weight that is nonzero only near that subdomain's center. The arrival time is the smooth $L^\infty$ (log-sum-exp) distance between the composite embeddings of the start and goal. This replaces output-space continuity constraints with a latent-space distance, so overlapping subdomains are stitched together by weighted averages of embeddings rather than by matching scalar values at boundary points.

What would settle it

Train FB-NTFields on a large map, then compare predicted $T(q_s,q_g)$ against the true shortest-path time from a fine-grid FMM for many start-goal pairs whose shortest paths cross multiple subdomain boundaries; if the prediction error is large or the gradient-following paths collide for a nontrivial fraction of pairs, the representational claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the Eikonal cost-to-go $T(q_s,q_g)$ can be represented as $T(q_s,q_g)=\alpha\log\sum\exp\left(\beta\,|\Phi_w(q_s)-\Phi_w(q_g)|\right)$, where $\Phi_w$ is a weighted composition of subdomain-specific latent encoders, and that training this representation end-to-end with the Eikonal speed loss yields an arrival-time field whose negative gradient guides bidirectional path search. The paper reports that this scheme outperforms NTFields and P-NTFields, which enforce continuity only at individual spatial points and therefore fail to capture the two-point spatial connectivity required for motion planning in large environments.

Load-bearing premise

The load-bearing assumption is that the learned latent distance in equation (5) can represent the true Eikonal travel time for every start-goal pair in a large obstacle-filled environment once the PDE loss has been minimized on randomly sampled pairs, even though no proof is given that this distance obeys metric properties or propagates correctly across subdomain boundaries.

Editorial extensions

If this is right

Planning in Gibson and real-world maps up to 1859.6 square meters completes in under 0.1 seconds on average with a success rate near 97 percent, against baselines that either collapse or take seconds.
Training requires only random start-goal samples and distance-to-obstacle queries, no expert trajectories, so the approach can be applied to newly mapped environments in minutes.
The latent-distance representation is symmetric by construction, so swapping start and goal leaves the predicted time field unchanged.
Bidirectional gradient descent with speed-scaled step sizes produces smooth paths that a Unitree B1 quadruped can follow indoors.
Because subdomains are small, the largest FB-NTFields model uses fewer parameters than the smallest NTFields and P-NTFields models tested, while converging in far fewer epochs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the latent-distance representation reliably approximates the cost-to-go, the same architecture could be adapted to other two-point prediction problems, such as learned distance fields or pairwise collision likelihood, where pointwise scalar fields are awkward.
The log-sum-exp form interpolates between max-norm and average-norm behavior in latent space, so tuning $\alpha$ and $\beta$ may trade off path safety against path directness, an effect the paper does not explore.
The Gaussian-weighted subdomain composition should allow incremental updates: when a map changes locally, only the affected subdomain encoders would need retraining, which the paper does not test.
A natural stress test would be to check whether the learned embedded distance satisfies the triangle inequality on random triples; if it approximately does, the latent space behaves as a learned metric space, which would clarify why gradient-descent path inference works.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know about this paper. First, it is a genuine empirical advance for physics-informed neural motion planning: FB-NTFields scales to large indoor maps (up to 1859 sq. m), achieves 97.4% success on Gibson environments where prior PiNMPs sit around 65%, and includes a real Unitree B1 demo with planning times around 0.05–0.15 s. That is a real result. Second, the theoretical grounding for the key architectural choice is thin, and the paper would benefit from an honest analysis of the terminal phase of path inference.

The new combination is overlapping subdomain encoders with a symmetric latent-space distance (LogSumExp of absolute latent differences) for the arrival time. The paper correctly cites NTFields for the symmetric operator and FBPINNs/KiloNeRF for domain decomposition. The ablation in Fig. 3 is convincing: larger single-MLP NTFields/P-NTFields do not close the gap, while adding subdomains does. The real-robot validation is meaningful evidence.

The soft spots are real but addressable. The stress-test concern about the diagonal is plausible: because the representation is symmetric and smooth, the gradient at an exact coincident start–goal is zero, which is inconsistent with the Eikonal boundarcondition. But the paper never samples exactly on the diagonal, and bidirectional descent stops before reaching it, so this may be a gap in analysis rather than a demonstrated failure. Still, the step-size scaling in Eq. 8 (γS²) could become unstable near the goal, and the paper gives no evidence about how close the descent comes to the singular region. I would ask the authors to report the distribution of final distances and the gradient magnitudes in the terminal phase.

Other issues: code and hyperparameters are only promised; success rates have no error bars; and the real-world comparison omits the PiNMP baselines because they did not converge, which is defensible but limits the claim. The central representational assumption—that a latent distance can represent a globally consistent cost-to-go—is unproven, but the empirical evidence gives some confidence.

This paper deserves a serious referee. I would recommend conditional acceptance: the empirical core is solid, and the theoretical gaps can be addressed in revision. For a researcher in learning-based motion planning, it is worth reading; I would likely cite it if I worked in this subfield.

Referee Report

3 major / 5 minor

Summary. The paper introduces FB-NTFields, a neural representation for the Eikonal cost-to-go in motion planning. The workspace is partitioned into overlapping subdomains, each assigned a SIREN encoder with Fourier features; a global encoding Phi_w(q) is formed by Gaussian-weighted averaging of subdomain encodings, and the arrival time is defined as T(q_s,q_g)=alpha*log(sum exp(beta*|Phi_w(q_s)-Phi_w(q_g)|)). Training uses a speed-ratio loss derived from the Eikonal equation against a TSDF-based speed field, and inference follows negative gradients of T bidirectionally. Experiments compare the method with NTFields, P-NTFields, RRT-Connect, LazyPRM, and FMM on ten Gibson environments and two large real-world maps, including a Unitree B1 deployment.

Significance. If the reported results are reproducible, FB-NTFields would be a meaningful advance: it achieves roughly 97% success rate on Gibson maps where NTFields and P-NTFields are around 65%, with planning times of 0.05-0.1 s, lower parameter counts, and no demonstration data. The real-robot deployment and the comparisons against classical planners and FMM ground truth are relevant empirical evidence. The central representational assumption - that a smooth latent distance of the proposed form can represent the Eikonal solution - is, however, not analyzed, and the paper's reproducibility is incomplete because hyperparameters, seeds, and code are not provided.

major comments (3)

[IV-D, Eq. (5); IV-F, Eq. (8)] The representation T(q_s,q_g)=alpha*log(sum exp(beta*|Phi_w(q_s)-Phi_w(q_g)|)) is symmetric in the two endpoints, and with the standard autodiff subgradient of |.| at 0 it satisfies nabla_{q_s}T(q,q)=0 and nabla_{q_g}T(q,q)=0 for every q. The true Eikonal arrival time has a non-smooth cusp at the source, with ||nabla T|| approaching 1/S*(q) as the moving endpoint approaches the source, not a smooth minimum. Because the training objective in Eq. (6) samples q_s and q_g independently, the near-diagonal region is essentially never penalized, so the predicted speed S=1/||nabla T|| can diverge there. In the bidirectional update of Eq. (8), the term gamma*S^2*nabla T then has magnitude gamma/||nabla T||, which is uncontrollable near the diagonal. The paper should add a boundary/diagonal loss term, sample close start-goal pairs, or report the sensitivity of the reported success rates to the stopping threshold and step size gamma.
[IV-C, Eq. (4); IV-D, Eq. (5)] The global coherence claim is not established. The Gaussian-weighted average of subdomain encodings guarantees differentiability of Phi_w with respect to q, but it does not by itself guarantee that the induced latent distance satisfies the triangle inequality, that arrival times propagate correctly across subdomain boundaries, or that gradients of T agree with the true geodesic direction when start and goal lie in different subdomains. Eq. (6) constrains only local gradient magnitudes at sampled pairs. I recommend an ablation that reports success rate and path length separately for start-goal pairs that cross subdomain boundaries, and a comparison of the learned T with FMM along such pairs.
[V-B, Table II; V-C, Table III; IV-G] The empirical core of the paper is not reproducible from the manuscript alone. The hyperparameters alpha, beta, gamma, subdomain count, overlap and Gaussian variance, and the speed-field constants d_min_obs, d_max_obs, and s_const are not specified; success rates are given as point estimates without seeds or confidence intervals; and the code is only promised for the final version. Since the main claim is the large success-rate improvement over NTFields and P-NTFields, the paper should include the configuration (or code) and report variance across at least three seeds, together with the stopping threshold used in Eq. (8).

minor comments (5)

[III] The sentence 'finding an obstacle-free trajectory tau subset Q_free comparing configuration sequence connecting...' appears to have a typo; 'comparing' should likely be 'comprising' or 'connecting'.
[III, Eq. (1)] Eq. (1) writes the Eikonal condition only for nabla_{q_g}T; since Eq. (6) also uses nabla_{q_s}T, please state the sign convention for the gradient with respect to the start.
[V-B, Table II] The text says the table demonstrates the success rate across each environment, but Table II reports aggregate values only; please report per-environment success rates or correct the sentence.
[IV-F] The stopping threshold for the bidirectional update is not specified; please provide it, as it is directly relevant to the terminal-phase behavior discussed in the major comments.
[IV-G] The implementation section says the authors plan to release code with the final version; please include a reproducibility statement with hyperparameters and environment splits in the manuscript.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the method is trained against an externally specified TSDF speed field and evaluated against classical planners; the only same-author citation is a non-load-bearing design credit.

full rationale

FB-NTFields is self-contained against external benchmarks. Equation (5) defines the arrival time as a LogSumExp over latent differences, and Equation (6) trains the predicted Eikonal speed S = 1/||grad T|| to match the TSDF-derived speed S*(q), which is an externally specified function of obstacle distance. The reported success rates in Tables II and III are measured on start-goal pairs through the resulting time field and are compared against RRTConnect, LazyPRM, and FMM on Gibson and real-world maps, so the planning claim is not equivalent to a fitted target. The only citation to the authors' own prior work is the sentence in Section IV-D crediting the symmetric operator to NTFields; this is an explicit architectural inspiration, and no argument in the paper depends on that prior paper for validity. No uniqueness theorem or prior result is imported to force the architecture. The diagonal-gradient issue noted by a skeptical reader is a soundness limitation of the symmetric ansatz, but it is not a circularity: it does not make any output equal to an input or reduce a prediction to a fit. Accordingly, the derivation chain is not circular, and the score reflects only the minor self-citation.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The method's central innovation is architectural; it introduces no new physical entity. The key representational assumption, that a latent-distance field coincides with the Eikonal arrival time, is introduced without proof. The main free choices are the distance scaling hyperparameters and the domain decomposition geometry, neither of which is reported in enough detail for replication.

free parameters (4)

alpha and beta scaling factors = not reported
Introduced in Eq. 5 as the scale of the smooth L-infinity latent distance. They control the magnitude of predicted arrival time and thus the path-inference dynamics, but their values are not given.
subdomain count, arrangement, overlap, and Gaussian variance = not reported
The manual domain decomposition geometry is a central design choice. The paper says subdomains are overlapping and Gaussian weights are used, but gives no per-environment values or sensitivity analysis.
speed-field constants d_min_obs, d_max_obs, s_const = not reported
These constants in Eq. 7 define the ground-truth speed field from the truncated signed distance function. They are chosen by hand and influence both training and path safety.
path-inference step size gamma and iteration count = not reported
Used in Eq. 8 for bidirectional gradient descent during path extraction. The values are not reported, yet they directly affect planning time, path quality, and success rate.

assumptions (6)

standard math The Eikonal PDE viscosity solution defines the true arrival time (cost-to-go) field from a start to a goal.
Invoked in Section III and used to design the loss in Eq. 6. This is a classical result, but the paper relies on it without revisiting the uniqueness conditions.
ad hoc to paper A function of the form of a smooth L-infinity distance between learned pointwise latent embeddings can represent the Eikonal solution for arbitrary start-goal pairs.
Introduced in Section IV-D, Eq. 5. This is the paper's core architectural assumption and is never derived or bounded. It is the most fragile premise.
ad hoc to paper Gaussian-weighted averaging of overlapping subdomain encodings preserves differentiability and global consistency across subdomain boundaries.
Stated in Section IV-C, Eq. 4. No analysis shows that the weighted composition guarantees the composite field satisfies the Eikonal equation or that boundary artifacts are absent.
domain assumption SIREN activations and Fourier features can mitigate spectral bias sufficiently within each subdomain.
Used in Section IV-B, based on prior work [34], [35]. This is a reasonable empirical assumption, but no local convergence evidence is provided in this paper.
domain assumption The speed field S* derived from a truncated signed distance to obstacles is an adequate ground-truth speed for the Eikonal equation in motion planning.
Defined in Eq. 7 and used as the training target. This is a standard modeling choice, but it implicitly equates distance-to-obstacle-based speed with the true optimal motion cost.
domain assumption Training on randomly sampled start-goal pairs generalizes to unseen pairs and to new environments.
The method is trained on sampled pairs and evaluated on other pairs and maps, but no generalization bound or coverage argument is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Physics-informed Neural Motion Planning via Domain Decomposition in Large Environments." pith.science (2026). https://pith.science/paper/7NFBJKYM

@misc{pith2026250612742,
  author       = {Pith},
  title        = {Pith review of: Physics-informed Neural Motion Planning via Domain Decomposition in Large Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7NFBJKYM}},
  note         = {Machine review of arXiv:2506.12742}
}

read the original abstract

Physics-informed Neural Motion Planners (PiNMPs) provide a data-efficient framework for solving the Eikonal Partial Differential Equation (PDE) and representing the cost-to-go function for motion planning. However, their scalability remains limited by spectral bias and the complex loss landscape of PDE-driven training. Domain decomposition mitigates these issues by dividing the environment into smaller subdomains, but existing methods enforce continuity only at individual spatial points. While effective for function approximation, these methods fail to capture the spatial connectivity required for motion planning, where the cost-to-go function depends on both the start and goal coordinates rather than a single query point. We propose Finite Basis Neural Time Fields (FB-NTFields), a novel neural field representation for scalable cost-to-go estimation. Instead of enforcing continuity in output space, FB-NTFields construct a latent space representation, computing the cost-to-go as a distance between the latent embeddings of start and goal coordinates. This enables global spatial coherence while integrating domain decomposition, ensuring efficient large-scale motion planning. We validate FB-NTFields in complex synthetic and real-world scenarios, demonstrating substantial improvements over existing PiNMPs. Finally, we deploy our method on a Unitree B1 quadruped robot, successfully navigating indoor environments. The supplementary videos can be found at https://youtu.be/OpRuCbLNOwM.

Figures

Figures reproduced from arXiv: 2506.12742 by the authors.

**Figure 2.** FB-NTFields workflow decomposes the given domain into overlapping subdomains. The given start, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Ablation comparison across three Gibson environ [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Depiction of two Gibson environments: The figure illustrates the paths generated by each method between a same [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Mollified Value Learning
cs.LG 2026-02 conditional novelty 5.0 of 10

Mollified Value Learning regularizes offline goal-conditioned value estimates with a Feynman-Kac expectation version of the viscous HJB equation instead of a pointwise Eikonal constraint.

Reference graph

Works this paper leans on

39 extracted references · 26 canonical work pages · cited by 1 Pith paper

[1]

Planning algorithms,

S. LaValle, “Planning algorithms,”Cambridge University Press google schola, vol. 2, pp. 3671–3678, 2006

work page 2006
[2]

Sampling-based algorithms for optimal motion planning,

S. Karaman and E. Frazzoli, “Sampling-based algorithms for optimal motion planning,”The international journal of robotics research, vol. 30, no. 7, pp. 846–894, 2011

2011
[3]

Rapidly-exploring ran- dom trees: Progress and prospects,

S. M. LaValle, J. J. Kuffner, B. Donaldet al., “Rapidly-exploring ran- dom trees: Progress and prospects,”Algorithmic and computational robotics: new directions, vol. 5, pp. 293–308, 2001

work page 2001
[4]

Informed RRT: Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic,

J. D. Gammell, S. S. Srinivasa, and T. D. Barfoot, “Informed RRT: Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic,” in2014 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2014, pp. 2997–3004

work page 2014
[5]

Fast marching tree: A fast marching sampling-based method for optimal motion planning in many dimensions,

L. Janson, E. Schmerling, A. Clark, and M. Pavone, “Fast marching tree: A fast marching sampling-based method for optimal motion planning in many dimensions,”The International journal of robotics research, vol. 34, no. 7, pp. 883–921, 2015

2015
[6]

Motion planning networks,

A. H. Qureshi, A. Simeonov, M. J. Bency, and M. C. Yip, “Motion planning networks,” in2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 2118–2124

2019
[7]

Motion planning networks: Bridging the gap between learning-based and classical motion planners,

A. H. Qureshi, Y . Miao, A. Simeonov, and M. C. Yip, “Motion planning networks: Bridging the gap between learning-based and classical motion planners,”IEEE Transactions on Robotics, vol. 37, no. 1, pp. 48–66, 2020

2020
[8]

Learning sampling dis- tributions for robot motion planning,

B. Ichter, J. Harrison, and M. Pavone, “Learning sampling dis- tributions for robot motion planning,” in2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2018, pp. 7087–7094

work page 2018

Show all 39 references

[9]

Deeply informed neural sampling for robot motion planning,

A. H. Qureshi and M. C. Yip, “Deeply informed neural sampling for robot motion planning,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2018, pp. 6582– 6588

2018
[10]

Lego: Leveraging experience in roadmap generation for sampling-based planning,

R. Kumar, A. Mandalika, S. Choudhury, and S. Srinivasa, “Lego: Leveraging experience in roadmap generation for sampling-based planning,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 1488–1495

2019
[11]

Differentiable spatial plan- ning using transformers,

D. S. Chaplot, D. Pathak, and J. Malik, “Differentiable spatial plan- ning using transformers,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 1484–1495

2021
[12]

NTFields: Neural time fields for physics- informed robot motion planning,

R. Ni and A. H. Qureshi, “NTFields: Neural time fields for physics- informed robot motion planning,” inInternational Conference on Learning Representations, 2023

2023
[13]

Progressive learning for physics-informed neural motion plan- ning,

——, “Progressive learning for physics-informed neural motion plan- ning,”arXiv preprint arXiv:2306.00616, 2023

2023 arXiv
[14]

Pc-planner: Physics-constrained self-supervised learning for robust neural motion planning with shape-aware distance function,

X. Shen, H. Peng, Z. Yang, J. Xu, H. Bao, R. Hu, and Z. Cui, “Pc-planner: Physics-constrained self-supervised learning for robust neural motion planning with shape-aware distance function,” in SIGGRAPH Asia 2024 Conference Papers, 2024, pp. 1–11

2024
[15]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational physics, vol. 378, pp. 686–707, 2019

2019
[16]

Physics-informed machine learning,

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics-informed machine learning,”Nature Reviews Physics, vol. 3, no. 6, pp. 422–440, 2021

2021
[17]

Viscosity solutions of hamilton- jacobi equations,

M. G. Crandall and P.-L. Lions, “Viscosity solutions of hamilton- jacobi equations,”Transactions of the American mathematical society, vol. 277, no. 1, pp. 1–42, 1983

1983
[18]

Frequency principle: Fourier analysis sheds light on deep neural networks,

Z.-Q. J. Xu, Y . Zhang, T. Luo, Y . Xiao, and Z. Ma, “Frequency principle: Fourier analysis sheds light on deep neural networks,”arXiv preprint arXiv:1901.06523, 2019

1901 arXiv
[19]

Characterizing possible failure modes in physics-informed neural networks,

A. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, and M. W. Mahoney, “Characterizing possible failure modes in physics-informed neural networks,”Advances in Neural Information Processing Systems, vol. 34, pp. 26 548–26 560, 2021

2021
[20]

Finite basis physics- informed neural networks (fbpinns): a scalable domain decomposition approach for solving differential equations,

B. Moseley, A. Markham, and T. Nissen-Meyer, “Finite basis physics- informed neural networks (fbpinns): a scalable domain decomposition approach for solving differential equations,”Advances in Computa- tional Mathematics, vol. 49, no. 4, p. 62, 2023

2023
[21]

Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps,

C. Reiser, S. Peng, Y . Liao, and A. Geiger, “Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 14 335–14 345

2021
[22]

A fast marching level set method for monotonically advancing fronts,

J. A. Sethian, “A fast marching level set method for monotonically advancing fronts,”Proceedings of the National Academy of Sciences, vol. 93, no. 4, pp. 1591–1595, 1996

1996
[23]

Direct and indirect methods for trajectory optimization,

O. V on Stryk and R. Bulirsch, “Direct and indirect methods for trajectory optimization,”Annals of operations research, vol. 37, pp. 357–373, 1992

1992
[24]

Path-constrained trajectory opti- mization using sparse sequential quadratic programming,

J. T. Betts and W. P. Huffman, “Path-constrained trajectory opti- mization using sparse sequential quadratic programming,”Journal of Guidance, Control, and Dynamics, vol. 16, no. 1, pp. 59–68, 1993

1993
[25]

Chomp: Covariant hamiltonian optimization for motion planning,

M. Zucker, N. Ratliff, A. D. Dragan, M. Pivtoraiko, M. Klingensmith, C. M. Dellin, J. A. Bagnell, and S. S. Srinivasa, “Chomp: Covariant hamiltonian optimization for motion planning,”The International journal of robotics research, vol. 32, no. 9-10, pp. 1164–1193, 2013

2013
[26]

Value iter- ation networks,

A. Tamar, Y . Wu, G. Thomas, S. Levine, and P. Abbeel, “Value iter- ation networks,”Advances in neural information processing systems, vol. 29, 2016

2016
[27]

Prm-rl: Long-range robotic navigation tasks by combining reinforcement learning and sampling-based planning,

A. Faust, K. Oslund, O. Ramirez, A. Francis, L. Tapia, M. Fiser, and J. Davidson, “Prm-rl: Long-range robotic navigation tasks by combining reinforcement learning and sampling-based planning,” in 2018 IEEE international conference on robotics and automation (ICRA). IEEE, 2018,...

2018
[28]

Universal planning networks: Learning generalizable representations for visuo- motor control,

A. Srinivas, A. Jabri, P. Abbeel, S. Levine, and C. Finn, “Universal planning networks: Learning generalizable representations for visuo- motor control,” inInternational Conference on Machine Learning. PMLR, 2018, pp. 4732–4741

2018
[29]

Control transformer: robot navigation in unknown environments through prm-guided return-conditioned sequence modeling,

D. Lawson and A. H. Qureshi, “Control transformer: robot navigation in unknown environments through prm-guided return-conditioned sequence modeling,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 9324–9331

2023
[30]

Physics-informed neural motion planning on constraint manifolds,

R. Ni and A. H. Qureshi, “Physics-informed neural motion planning on constraint manifolds,”arXiv preprint arXiv:2403.05765, 2024

2024 arXiv
[31]

Tensorf: Tensorial ra- diance fields,

A. Chen, Z. Xu, A. Geiger, J. Yu, and H. Su, “Tensorf: Tensorial ra- diance fields,” inEuropean conference on computer vision. Springer, 2022, pp. 333–350

2022
[32]

Instant neural graphics primitives with a multiresolution hash encoding,

T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”ACM transactions on graphics (TOG), vol. 41, no. 4, pp. 1–15, 2022

2022
[33]

Nerf: Representing scenes as neural radiance fields for view synthesis,

B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ra- mamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

2021
[34]

Fourier features let networks learn high frequency functions in low dimen- sional domains,

M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Ragha- van, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng, “Fourier features let networks learn high frequency functions in low dimen- sional domains,”Advances in Neural Information Processing Systems, vol. 33, ...

2020
[35]

Implicit neural representations with periodic activation functions,

V . Sitzmann, J. N. Martel, A. W. Bergman, D. B. Lindell, and G. Wetzstein, “Implicit neural representations with periodic activation functions,” inProc. NeurIPS, 2020

2020
[36]

igibson 2.0: Object-centric simulation for robot learning of everyday house- hold tasks,

C. Li, F. Xia, R. Mart ´ın-Mart´ın, M. Lingelbach, S. Srivastava, B. Shen, K. Vainio, C. Gokmen, G. Dharan, T. Jainet al., “igibson 2.0: Object-centric simulation for robot learning of everyday house- hold tasks,”arXiv preprint arXiv:2108.03272, 2021

2021 arXiv
[37]

RRT-connect: An efficient approach to single-query path planning,

J. J. Kuffner and S. M. LaValle, “RRT-connect: An efficient approach to single-query path planning,” inProceedings 2000 ICRA. Millen- nium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No. 00CH37065), vol. 2. IEEE, 2000, pp. 995–1001

2000
[38]

Path planning using lazy prm,

R. Bohlin and L. E. Kavraki, “Path planning using lazy prm,” in Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), vol. 1. IEEE, 2000, pp. 521–528

2000
[39]

GEASI: Geodesic-based earliest activation sites identification in cardiac models,

T. Grandits, A. Effland, T. Pock, R. Krause, G. Plank, and S. Pezzuto, “GEASI: Geodesic-based earliest activation sites identification in cardiac models,”International Journal for Numerical Methods in Biomedical Engineering, vol. 37, no. 8, p. e3505, 2021. [Online]. Available:...

2021 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Planning algorithms,

S. LaValle, “Planning algorithms,”Cambridge University Press google schola, vol. 2, pp. 3671–3678, 2006

work page 2006

[2] [2]

Sampling-based algorithms for optimal motion planning,

S. Karaman and E. Frazzoli, “Sampling-based algorithms for optimal motion planning,”The international journal of robotics research, vol. 30, no. 7, pp. 846–894, 2011

2011

[3] [3]

Rapidly-exploring ran- dom trees: Progress and prospects,

S. M. LaValle, J. J. Kuffner, B. Donaldet al., “Rapidly-exploring ran- dom trees: Progress and prospects,”Algorithmic and computational robotics: new directions, vol. 5, pp. 293–308, 2001

work page 2001

[4] [4]

Informed RRT: Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic,

J. D. Gammell, S. S. Srinivasa, and T. D. Barfoot, “Informed RRT: Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic,” in2014 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2014, pp. 2997–3004

work page 2014

[5] [5]

Fast marching tree: A fast marching sampling-based method for optimal motion planning in many dimensions,

L. Janson, E. Schmerling, A. Clark, and M. Pavone, “Fast marching tree: A fast marching sampling-based method for optimal motion planning in many dimensions,”The International journal of robotics research, vol. 34, no. 7, pp. 883–921, 2015

2015

[6] [6]

Motion planning networks,

A. H. Qureshi, A. Simeonov, M. J. Bency, and M. C. Yip, “Motion planning networks,” in2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 2118–2124

2019

[7] [7]

Motion planning networks: Bridging the gap between learning-based and classical motion planners,

A. H. Qureshi, Y . Miao, A. Simeonov, and M. C. Yip, “Motion planning networks: Bridging the gap between learning-based and classical motion planners,”IEEE Transactions on Robotics, vol. 37, no. 1, pp. 48–66, 2020

2020

[8] [8]

Learning sampling dis- tributions for robot motion planning,

B. Ichter, J. Harrison, and M. Pavone, “Learning sampling dis- tributions for robot motion planning,” in2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2018, pp. 7087–7094

work page 2018