Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Transformer-based Heuristic for Advanced Air Mobility Planning

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A transformer trained on solver-generated heuristics can speed safety-constrained A* search for drone planning by up to 39.5% while keeping paths near-optimal.

desk verdict An incremental but honest extension of the authors' own learned-heuristic work; the main acceleration claim hinges on an unresolved timing ambiguity and an under-specified oracle. read the letter →

arxiv 2411.14427 v1 pith:RIMF6RQV submitted 2024-11-21 cs.RO

Jun Xiang , Jun Chen This is my paper · ORCID
classification cs.RO
keywords ConstrainedShortestPathA*algorithmAdvancedAirMobilityPlanningSupervisedLearningTransformerheuristicsearchrisk-awareunmannedaerialvehicles
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

This paper tries to establish that a transformer can learn the search heuristic that makes A*-style planners fast on risk-constrained drone routing. The problem is hard in the worst case: constrained shortest path (CSP) is NP-hard, and exact solvers are slow. The authors feed their ASD A* algorithm—A* augmented with a safety dimension—heuristics generated by two transformer networks, one that scores every grid cell given start and goal, and one that scores only the current search node. On random maps up to 32×32, the first network's heuristic beats the Manhattan baseline by exploring 8.71–39.5% fewer nodes and searching 9.71–24% faster while keeping success-weighted path length (SPL) above 97.55%. If true, this means learning can supply the guidance that makes constrained planning practical for urban air mobility without surrendering the planner's validity guarantee.

What carries the argument

The load-bearing mechanism is the Transformer encoder used as a heuristic generator. For Riskmap2.0, the flattened risk map and tokenized start/destination are embedded and combined, then processed by the encoder; the output layer classifies each grid's heuristic into a discrete value set and takes an argmax, which keeps the generated heuristics within a bounded range. For Riskmap-state, the risk map is padded to a fixed size, combined with sinusoidal positional embeddings, concatenated with a task embedding for the current node and destination, and processed by the same encoder, with an MSE loss on the single scalar output. The training labels are constructed so that grids on the solver's shortest path receive lower heuristic values (Manhattan distance), while off-path grids receive a penalty; an accurate learned version of that pattern is what lets ASD A* concentrate its expansion on the shortest path.

What would settle it

Run ASD A* with the Riskmap2.0 heuristic on maps whose risk distribution differs from the equal-thirds training distribution (for example, mostly safe cells with a few high-risk corridors) and compare node expansions against the Manhattan heuristic; if the learned heuristic does not reduce expansions or SPL drops materially below the Manhattan baseline on any such map family, the generalization claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a transformer trained on solver-generated heuristic labels can produce, at inference time, a heuristic that guides ASD A* to near-shortest feasible paths while expanding far fewer nodes than the Manhattan-distance baseline. The Riskmap2.0 network takes a risk map, start, and destination and outputs a heuristic value for every grid, trained with a classification loss over a bounded set of heuristic values; the Riskmap-state network takes the risk map, the current search node, and the destination and regresses a single heuristic. The reported results put Riskmap2.0's node-exploration reduction at 8.71–39.5% across map sizes with SPL at 97.55–99.58%, and on a 16×16 city wind-flow map the reduction reaches 53.53% with SPL 99.73%. The paper presents these numbers as evidence that learned heuristics can accelerate safety-constrained air mobility planning while preserving near-optimal path quality.

Load-bearing premise

The speedup claim depends on the unreported traditional solver's labels and the hand-chosen Manhattan-distance penalty being a reliable oracle for the planner, and on the transformer generalizing from synthetic maps with equal thirds of safe, low-risk, and high-risk grids to the deployment environment.

Editorial extensions

If this is right

  • On random 16×16, 24×24, and 32×32 risk maps, ASD A* with the Riskmap2.0 heuristic explores 8.71–39.5% fewer nodes and searches 9.71–24% faster than with Manhattan distance, while SPL stays between 97.55% and 99.58%.
  • On a structured 16×16 city wind-flow map, the same heuristic explores 53.53% fewer nodes and searches 48.65% faster, with SPL 99.73%, showing that the learned heuristic can transfer to at least one non-random risk distribution.
  • The Riskmap-state network, trained jointly on 16×16 and 64×64 maps, produces per-node heuristics with MSE 1.742 on 16×16 and 6.234 on 64×64, reducing node exploration by 19.23% and 10.51%, respectively.
  • Because the transformer only proposes heuristics and ASD A* remains the planner, every returned path is still checked against the safety constraint—unlike end-to-end transformer planners that can fail to produce valid plans.

Reading between the lines

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

  • The reported advantage shrinks as map size grows (39.5% at 16×16, 16.75% at 24×24, 8.71% at 32×32); the authors attribute this to reduced training data per map, which suggests the larger-map gap could close with more labels or longer training rather than a fundamental limit of the heuristic.
  • The method's ceiling is set by the undisclosed solver that generates the labels; if that solver is approximate, the learned heuristic inherits its errors, and an exact constrained-shortest-path solver would provide a stronger oracle and an upper bound on achievable speedup.
  • The only out-of-distribution test is a 16×16 city map, so claims of adaptability to different map sizes and risk structures would be much stronger if repeated at 64×64 or on real wind-flow maps without downsampling.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes two transformer-based heuristic generators for the ASD A* algorithm, which solves a risk-constrained shortest path (CSP) problem for Advanced Air Mobility planning. The first generator, Riskmap2.0, takes a risk map, start, and destination and outputs a heuristic for every grid; the second, Riskmap-state, takes the current node, destination, and risk map and outputs a heuristic for that node. The heuristics are trained in a supervised manner on labels produced by an unspecified traditional solver. Experiments on 16x16, 24x24, and 32x32 random maps and on a 16x16 city map report that Riskmap2.0 reduces node exploration and search time relative to a Manhattan baseline while maintaining SPL above 97.55%, and that Riskmap-state produces heuristics close to expert labels with smaller node counts. The central claim is that the learned heuristics accelerate ASD A* while preserving near-optimal path length.

Significance. If the reported results hold, the paper would demonstrate a practical way to inject learned, risk-aware guidance into a constrained A* planner, with a useful cross-distribution test on a city wind-flow map. Strengths include the large dataset sizes (over 128,000 entries per map size), the two distinct heuristic-generation strategies, and the explicit use of a classification-style output for Riskmap2.0 that keeps heuristics within a bounded value set. However, the central acceleration claim is currently not fully supported because the paper does not clarify whether the reported search times include transformer inference, does not disclose the expert-label generation procedure (solver and penalty value), and provides no variance or statistical tests for any of the reported averages. The comparison against Manhattan alone also leaves open how much of the speedup comes from the oracle-like construction of the training labels rather than from learned generalization. These issues are load-bearing for the paper's motivating claim of computational acceleration, but they are addressable with additional experiments and reporting.

major comments (4)
  1. [V.A, Table I] The paper reports 'Search time (ms)' for Riskmap2.0 and Manhattan, but it never states whether this is A*-only node-expansion time or end-to-end time including heuristic generation. Riskmap2.0 requires a transformer forward pass over the entire flattened map before A* begins; for a 32x32 map this is a 1024-token sequence, and the reported saving over Manhattan is only 2.64 ms (27.19 vs 24.55 ms). A single transformer inference can plausibly exceed that saving on commodity hardware, so if the reported times exclude inference, the abstract's 'accelerate' claim is not supported end-to-end. The authors must specify the timing protocol and report wall-clock time including all heuristic-generation costs for both the learned and baseline heuristics.
  2. [III.B.1] The expert heuristic for Riskmap2.0 is defined using an unnamed 'traditional solver' and an undisclosed penalty added to the Manhattan distance for grids not on the shortest path. This makes the dataset irreproducible and leaves a free parameter that directly controls how aggressively ASD A* is steered toward the oracle path. The manuscript also asserts that this heuristic 'maintains consistency' without proof; if the penalty is positive, the heuristic can violate consistency because an off-path node can have h(n) = Manhattan(n) + P while a neighboring on-path node has h(n') = Manhattan(n) - 1, giving h(n) > 1 + h(n'). The authors must report the solver, the penalty value, and either prove or check consistency and admissibility, since these properties affect both optimality (SPL) and node counts.
  3. [V.A, V.C] All reported improvements are averages over 1000 tasks, but no standard deviations, confidence intervals, or statistical tests are given. Several differences in Table I are small in absolute terms, such as the 32x32 node count (826.14 vs 754.19) and search time (27.19 vs 24.55 ms), and without variance information it is impossible to tell whether these differences are robust across seeds, tie-breaking rules, or task draws. The paper should include error bars and, where relevant, paired tests over the same task set.
  4. [III.B.2, V.B] For Riskmap-state the authors acknowledge that 'the heuristic generating time may be longer than the search time, because the Riskmap-state must generate a heuristic for every node A* found.' The same caveat applies to Riskmap2.0, where the heuristic is generated once per task, but no attempt is made to measure this overhead. Since the stated purpose of the method is acceleration, the absence of any end-to-end timing comparison for either model leaves the core practical claim unverified. A concrete test would be to report total planning time including all neural network forward passes and compare it with the Manhattan heuristic on the same hardware.
minor comments (5)
  1. [Table I] The label 'Manhanton' is a typo and should read 'Manhattan'.
  2. [Introduction] The phrase 'as less nodes as possible' should be 'as few nodes as possible' for grammatical correctness.
  3. [V.B] The sentence describing the 16x16 Riskmap-state result is duplicated: 'For the 16*16 dataset, the number of nodes explored is 19.23% less than the Manhattan. for the 16*16 dataset, the number of nodes explored is 10.51% less than the Manhattan.' The second sentence likely refers to a different map size or condition and should be corrected.
  4. [II, Figure 2] The text says 'Fig 2 shows an example of the ADS A*' but the algorithm is consistently called ASD A* elsewhere; please fix the acronym.
  5. [III.B.1] The phrase 'while maintaining consistency' is a substantive algorithmic claim, not a formatting issue; if the authors keep it, they should provide a proof or a citation. If they remove it, the sentence remains complete without it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned-heuristic evaluation is a standard held-out generalization test with an external city-map distribution shift.

full rationale

The derivation chain is a supervised-learning pipeline: a traditional solver produces expert heuristics; these labels train transformer networks; the networks' outputs are fed to ASD A* and compared against a Manhattan heuristic on held-out maps. None of these steps reduces to its own input by construction. The training labels do encode the solver's solution (Section III.B.1 says the Riskmap2.0 dataset 'basically contains the solution for the task'), and the expert heuristic is deliberately designed so that ASD A* expands only shortest-path nodes, making the oracle's advantage over Manhattan a design property rather than a discovered result. However, the paper's actual claim is about the learned networks: they must generalize to 100 unseen maps per size and to a 16x16 city-map distribution (Section V.C), and the reported SPL below 100% shows the learned heuristics are imperfect. The city-map test is an external distribution shift that provides independent grounding. The remaining concerns are non-circular: the 'traditional solver' and the off-path penalty are unreported, and 'search time' in Table I is not stated to include the transformer forward pass; the paper itself acknowledges the analogous overhead for Riskmap-state in Section V.B. These are reproducibility and measurement-validity issues, not circularity. Self-citation [26] introduces ASD A*, but the current experiments implement and evaluate the algorithm against a standard baseline, so the self-citation is not load-bearing to the circularity analysis.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces two neural network architectures but no new physical or conceptual entities. The central claim relies on a set of domain assumptions about risk maps, the existence of an exact solver, and an unspecified oracle construction penalty. The free parameters are not reported precisely enough to be reproduced.

free parameters (2)
  • off-path penalty in expert heuristic
    Section III.B.1 states that non-shortest-path grids get Manhattan distance plus a penalty, but the penalty value is never reported. This parameter shapes the training labels and directly influences the node-exploration reduction observed at test time.
  • random safety level range for Riskmap-state current nodes = 0.9 to 1
    Section III.B.2 says a random safety level between 0.9 and 1 is assigned to current nodes, but the distribution and discretization are not specified. This choice affects the training distribution of the regression model.
assumptions (4)
  • domain assumption There exists an exact 'traditional solver' for the CSP on the generated maps, and it is used to create expert heuristics.
    Section III.B assumes this solver without naming it; the correctness of the heuristic labels depends on its correctness.
  • domain assumption Risk values are independent per grid and accumulated safety is the product of node safety values.
    Equation (3) in Section II formalizes safety as a product over visited nodes, an assumption about how risks combine in real airspace.
  • domain assumption The random map generator produces maps with equal numbers of safe, low-risk, and high-risk cells, and the test distribution is the same.
    Section III.A states this map generation rule; all main experiments use this distribution, with one additional city-map test.
  • ad hoc to paper ASD A* can represent the safety dimension with a finite discretization that does not make the state space unmanageable.
    The paper never specifies the safety-level discretization, yet all experiments assume the added dimension is tractable. This is an unstated implementation assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transformer-based Heuristic for Advanced Air Mobility Planning." pith.science (2026). https://pith.science/paper/RIMF6RQV

@misc{pith2026241114427,
  author       = {Pith},
  title        = {Pith review of: Transformer-based Heuristic for Advanced Air Mobility Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RIMF6RQV}},
  note         = {Machine review of arXiv:2411.14427}
}
read the original abstract

Safety is extremely important for urban flights of autonomous Unmanned Aerial Vehicles (UAVs). Risk-aware path planning is one of the most effective methods to guarantee the safety of UAVs. This type of planning can be represented as a Constrained Shortest Path (CSP) problem, which seeks to find the shortest route that meets a predefined safety constraint. Solving CSP problems is NP-hard, presenting significant computational challenges. Although traditional methods can accurately solve CSP problems, they tend to be very slow. Previously, we introduced an additional safety dimension to the traditional A* algorithm, known as ASD A*, to effectively handle Constrained Shortest Path (CSP) problems. Then, we developed a custom learning-based heuristic using transformer-based neural networks, which significantly reduced computational load and enhanced the performance of the ASD A* algorithm. In this paper, we expand our dataset to include more risk maps and tasks, improve the proposed model, and increase its performance. We also introduce a new heuristic strategy and a novel neural network, which enhance the overall effectiveness of our approach.

Figures

Figures reproduced from arXiv: 2411.14427 by the authors.

Figure 1
Figure 1. An example of heuristic generators: a) Riskmap2.0 takes the given start position (orange letter S), destination (green [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. CSP example(2*2): The risk score of the grid (0, 0) is 0.1, the risk score of the grid (0, 1) and the grid (1, 1) is 0.05. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Proposed transformer heuristic architecture [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Concatenated embedding of the Riskmap-state(64*64) [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Result example(32*32 random): blue number means how many times a node is explored on this grid [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: City risk map with wind flow near building [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Result example (16*16 city): blue number means how many times a node is explored on this grid [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 22 canonical work pages

  1. [26]

    Learning-accelerated a* search for risk-aware path planning

    Jun Xiang, Junfei Xie, and Jun Chen. Learning-accelerated a* search for risk-aware path planning. In AIAA SCITECH 2024 Forum , page 2895, 2024

    work page 2024

  2. [1]

    Safety enhancement for deep reinforcement learning in autonomous separation assurance

    Wei Guo, Marc Brittain, and Peng Wei. Safety enhancement for deep reinforcement learning in autonomous separation assurance. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 348–354. IEEE, 2021

    work page 2021

  3. [2]

    Minimum-risk path planning for long-range and low-altitude flights of autonomous unmanned aircraft

    Simon Schopferer and Sebastian Benders. Minimum-risk path planning for long-range and low-altitude flights of autonomous unmanned aircraft. In AIAA Scitech 2020 Forum , page 0137, 2020

    work page 2020

  4. [3]

    Quality-of-service routing for sup- porting multimedia applications

    Zheng Wang and Jon Crowcroft. Quality-of-service routing for sup- porting multimedia applications. IEEE Journal on selected areas in communications, 14(7):1228–1234, 1996

    work page 1996

  5. [4]

    A generalized permanent labelling algorithm for the shortest path problem with time windows

    Martin Desrochers and Franc ¸ois Soumis. A generalized permanent labelling algorithm for the shortest path problem with time windows. INFOR: Information Systems and Operational Research, 26(3):191–212, 1988

    work page 1988

  6. [5]

    A hybrid particle swarm optimization–variable neighborhood search algorithm for constrained shortest path problems

    Yannis Marinakis, Athanasios Migdalas, and Angelo Sifaleras. A hybrid particle swarm optimization–variable neighborhood search algorithm for constrained shortest path problems. European Journal of Operational Research, 261(3):819–834, 2017

    work page 2017

  7. [6]

    Grasp: Greedy randomized adaptive search procedures

    Mauricio GC Resende and Celso C Ribeiro. Grasp: Greedy randomized adaptive search procedures. In Search methodologies, pages 287–312. Springer, 2014

    work page 2014

  8. [7]

    Asymmetry matters: Dynamic half-way points in bidirectional labeling for solving shortest path problems with resource constraints faster

    Christian Tilk, Ann-Kathrin Rothenb ¨acher, Timo Gschwind, and Stefan Irnich. Asymmetry matters: Dynamic half-way points in bidirectional labeling for solving shortest path problems with resource constraints faster. European Journal of Operational Research , 261(2):530–539, 2017

    work page 2017

Show all 29 references
  1. [8]

    Probabilistically guaranteed path planning for safe urban air mobility using chance constrained rrt

    Pengcheng Wu, Lin Li, Junfei Xie, and Jun Chen. Probabilistically guaranteed path planning for safe urban air mobility using chance constrained rrt. In AIAA Aviation 2020 Forum , page 2914, 2020

    2020

  2. [9]

    Modeling car-following behavior in heterogeneous traffic mixing human-driven, automated and connected vehicles: considering multitype vehicle interactions

    Ziyu Song and Haitao Ding. Modeling car-following behavior in heterogeneous traffic mixing human-driven, automated and connected vehicles: considering multitype vehicle interactions. Nonlinear Dynam- ics, 111(12):11115–11134, 2023

    2023

  3. [10]

    A dual algorithm for the constrained shortest path problem

    Gabriel Y Handler and Israel Zang. A dual algorithm for the constrained shortest path problem. Networks, 10(4):293–309, 1980

    1980

  4. [11]

    Improved preprocessing, labeling and scaling algorithms for the weight-constrained shortest path problem

    Irina Dumitrescu and Natashia Boland. Improved preprocessing, labeling and scaling algorithms for the weight-constrained shortest path problem. Networks: An International Journal , 42(3):135–153, 2003

    2003

  5. [12]

    Finding the k shortest paths

    David Eppstein. Finding the k shortest paths. SIAM Journal on computing, 28(2):652–673, 1998

    1998

  6. [13]

    Hybrid multiscale search for dynamic planning of multi-agent drone traffic

    Jun Xiang, Jun Chen, and Yanchao Liu. Hybrid multiscale search for dynamic planning of multi-agent drone traffic. Journal of Guidance, Control, and Dynamics , 46(10):1963–1974, 2023

    1963

  7. [14]

    The constrained shortest path problem: algorithmic ap- proaches and an algebraic study with generalization

    Ying Xiao, Krishnaiyan Thulasiraman, Guoliang Xue, Alp ´ar J¨uttner, and S Arumugam. The constrained shortest path problem: algorithmic ap- proaches and an algebraic study with generalization. AKCE International Journal of Graphs and Combinatorics , 2(2):63–86, 2005

    2005

  8. [15]

    Incremental sampling-based algorithms for optimal motion planning

    Sertac Karaman and Emilio Frazzoli. Incremental sampling-based algorithms for optimal motion planning. 2011

    2011

  9. [16]

    The focussed dˆ* algorithm for real-time replanning

    Anthony Stentz et al. The focussed dˆ* algorithm for real-time replanning. In IJCAI, volume 95, pages 1652–1659, 1995

    1995

  10. [17]

    Anytime dynamic a*: An anytime, replanning algorithm

    Maxim Likhachev, David I Ferguson, Geoffrey J Gordon, Anthony Stentz, and Sebastian Thrun. Anytime dynamic a*: An anytime, replanning algorithm. In ICAPS, volume 5, pages 262–271, 2005

    2005

  11. [18]

    Optimal any-angle pathfinding in practice

    Daniel Damir Harabor, Alban Grastien, Dindar ¨Oz, and Vural Aksakalli. Optimal any-angle pathfinding in practice. Journal of Artificial Intelli- gence Research, 56:89–118, 2016

    2016

  12. [19]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  13. [20]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weis- senborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arX...

    2010 arXiv

  14. [21]

    Bridging the domain gap: Self-supervised 3d scene understanding with foundation models

    Zhimin Chen, Longlong Jing, Yingwei Li, and Bing Li. Bridging the domain gap: Self-supervised 3d scene understanding with foundation models. Advances in Neural Information Processing Systems , 36, 2024

    2024

  15. [22]

    Planning- oriented autonomous driving

    Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning- oriented autonomous driving. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 17853–17862, 2023

    2023

  16. [23]

    Deep learning applications in the medical image recogni- tion

    Song Yukun. Deep learning applications in the medical image recogni- tion. American Journal of Computer Science and Technology , 2(2):22– 26, 2019

    2019

  17. [24]

    Beyond a*: Better planning with transformers via search dynamics bootstrapping

    Lucas Lehnert, Sainbayar Sukhbaatar, Paul Mcvay, Michael Rabbat, and Yuandong Tian. Beyond a*: Better planning with transformers via search dynamics bootstrapping. arXiv preprint arXiv:2402.14083 , 2024

    2024 arXiv

  18. [25]

    A generalist agent

    Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022

    2022 arXiv

  19. [27]

    On evaluation of embodied navigation agents

    Peter Anderson, Angel Chang, Devendra Singh Chaplot, Alexey Doso- vitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, et al. On evaluation of embodied navigation agents. arXiv preprint arXiv:1807.06757 , 2018

    2018 arXiv

  20. [28]

    Safe path planning of uav based on reinforcement learning in probabilistic environments

    James Gault, Jun Xiang, and Jun Chen. Safe path planning of uav based on reinforcement learning in probabilistic environments. In AIAA SCITECH 2023 Forum, page 2677, 2023

    2023

  21. [29]

    Solving the rubik’s cube without human knowledge

    Stephen McAleer, Forest Agostinelli, Alexander Shmakov, and Pierre Baldi. Solving the rubik’s cube without human knowledge. arXiv preprint arXiv:1805.07470, 2018

    2018 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.