Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Learning-Augmented Model-Based Multi-Robot Planning for Time-Critical Search and Inspection Under Uncertainty

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that feeding a neural network's risk estimates into an expected-cost multi-robot planner cuts inspection cost by 16-40 percent relative to greedy and nearest-location baselines.

desk verdict A coherent learning-augmented planner for multi-robot inspection with consistent simulated gains, but the headline claim about noisy sensor data is untested: the GNN is fed exact hazard locations, not noisy measurements. read the letter →

arxiv 2507.06129 v1 pith:HGRDSNIM submitted 2025-07-08 cs.RO cs.MA

Abhish Khanal , Joseph Prince Mathew , Cameron Nowzari , Gregory J. Stein This is my paper · ORCID
classification cs.ROcs.MA
keywords multi-robotplanninggraphneuralnetworkinspectionunderuncertaintytime-criticalsearchexpected-costMDPrecedinghorizondisasterresponseUAV
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

Time-critical inspection usually pits two goals against each other: race to the locations most likely to be damaged, and avoid wasting travel time on nearby low-risk sites. This paper tries to show that the best resolution is to let a graph neural network produce a per-location damage likelihood from noisy sensor data, then hand those likelihoods to a model-based multi-robot planner that minimizes expected cost. The planner treats the problem as an MDP in which every second costs an amount proportional to the total likelihood mass over uninspected points, so routing decisions naturally balance risk against distance. In procedural simulations with 12-36 points of interest and 1-5 robots, the authors report 16-40 percent lower average cost than learned-greedy and optimistic nearest-location baselines, and a two-quadcopter field test confirms the pattern. If correct, the approach offers a modular alternative to end-to-end reinforcement learning for disaster response and surveillance.

What carries the argument

The load-bearing machinery is the expected-cost MDP recursion of Eq. (4), with temporally extended joint actions that end when the first robot finishes its assigned PoI. The transition time is the minimum over robots of travel time plus inspection time, and the cost accrued during that transition is the per-unit-time penalty $K$ times the elapsed time times the total likelihood mass $\sum_{l\in L}P(l)$ over still-uninspected PoIs. This turns the POMDP into an MDP whose only learned parameters are the fixed likelihoods $P(l)$, supplied by a TransformerConv graph neural network operating on a graph of PoIs and wind pockets with one-hot class features and distance-based edge features. A receding-horizon heuristic keeps the factorial action space manageable by selecting the six most-likely and six nearest PoIs, and a depth-first branch-and-bound search solves the recursion.

What would settle it

A reader could rerun the paper's 12-PoI, 3-robot experiment in a clustered-damage environment, giving a baseline planner the same GNN predictions but updating $P(l)$ by Bayes' rule after each inspection; if that baseline's average cost falls below the reported 221.1 for the model-based planner by more than trial noise, the fixed-likelihood MDP reduction in Eq. (4) is the limiting assumption rather than the learned estimates.

Watch

Extended reading notes

Core claim

The paper's central claim is that time-critical multi-robot inspection under uncertainty, despite being a POMDP in general, is well served by treating it as an MDP with expected-cost state-action values. For a state $s_t=\{L,q_t\}$ of uninspected PoIs and robot poses, and a joint action $a_t$ assigning each robot a distinct PoI, the first completed inspection occurs after $T(s_t,a_t)=\min_{l\in a_t,q\in q_t}[T_{\mathrm{travel}}(q\to l)+T_R(l)]$, and the recursion $$Q^*(s_t,a_t)=K T(s_t,a_t)\sum_{l\in L}P(l)+\min_{a_{t+1}\in A(s_{t+1})} Q^*(\{L\setminus\{L(s_t,a_t)\},q_{t+1}\},a_{t+1})$$ accumulates expected cost. When the fixed likelihoods $P(l)$ come from a graph neural network over points of interest and noisy wind-pocket observations, the resulting planner outperforms both a nearest-location optimistic planner and a learned greedy planner: at least 16.3%, 26.7%, and 26.2% lower average cost for 1, 3, and 5 robots in 12-PoI environments, 34.0%-39.9% lower in 24- and 36-PoI environments, and a 20.43% saving in a two-quadcopter field trial.

Load-bearing premise

The plan assumes each location's damage likelihood $P(l)$ is fixed, known from the GNN, and independent of what the robots observe during the mission, so inspecting one site never changes the estimated risk of another; if real damage is correlated or sensor estimates are biased, the fixed-probability plan can fall well short of optimal.

Editorial extensions

If this is right

  • In 12-PoI environments, the learned model-based planner lowers average cost by at least 16.3%, 26.7%, and 26.2% for 1, 3, and 5 robots compared with both baselines.
  • In 24- and 36-PoI environments, the reported improvement over baselines reaches 34.0%, 39.9%, and 37.3% for 1, 3, and 5 robots, respectively, suggesting the method scales rather than degrading.
  • Representative trajectory plots show that the planner finds damaged PoIs in less distance traveled, keeping the cumulative cost curve below both baselines throughout each trial.
  • A two-quadcopter field experiment with 10 PoIs yields a 20.43% cost saving over the optimistic planner, supporting real-world deployability.
  • Because the learned component outputs only per-location likelihoods, the same trained model can inform planning for different team sizes without retraining robot-specific policies.

Reading between the lines

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

  • The fixed-likelihood recursion in Eq. (4) never updates $P(l)$ after inspecting a location; in spatially clustered damage, a planner that re-estimates probabilities after each finding should outperform it, so the reported gains are plausibly a lower bound for a belief-updating extension.
  • The simulation gives the GNN exact wind-pocket node locations, so the field advantage may shrink when wind pockets must be inferred from noisy radar; testing with observation noise added to sensor features would reveal how much of the gain comes from clean inputs.
  • The same learning-plus-MDP decomposition could be applied to other time-critical monitoring tasks, but the paper's experiments only cover the storm-damage scenario, so transfer would need new sensor features and cost calibrations.
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 / 4 minor

Summary. The paper proposes a learning-augmented multi-robot planner for time-critical inspection of Points of Interest (PoIs) under uncertainty. A graph neural network estimates the probability that each PoI requires immediate attention, and a model-based planner using the Bellman recursion in Eq. (4) computes joint robot actions that minimize expected cost. The authors evaluate the planner in procedural storm-damage scenarios with 12, 24, and 36 PoIs and up to 5 robots, comparing against an optimistic nearest-location baseline and a learned greedy baseline, and they report percentage cost improvements over both. A two-quadcopter demonstration is also described.

Significance. If the claimed gains are robust, the paper offers a useful architectural contribution: it combines learned likelihood estimates with model-based multi-robot planning and makes the computation tractable via receding-horizon branch-and-bound search. The formal Bellman recursion is clearly stated, and the 500-trial simulation protocol is a solid start. However, the current evidence does not yet support the headline claim. The experiments do not exercise the 'noisy sensor data' pathway described in the abstract, and the mean-only statistics leave the magnitude and reliability of the improvements unclear. The contribution is potentially significant, but additional validation is needed before the central claims can be accepted.

major comments (4)
  1. [Section IV-B, Section IV-C] The evaluation does not exercise the 'noisy sensor data' pathway that is central to the paper's claim. In Section IV-B, the GNN input graph contains nodes for the same high-wind pockets used to generate ground-truth damage via Eq. (5), with one-hot class features and distance-based edge features; no experiment feeds the planner or the GNN with actual noisy sensor measurements, uncertain pocket locations, or missing/erroneous detections. Consequently, the reported improvements are obtained with near-ground-truth likelihoods and may not persist under realistic observation noise. Please add experiments with noisy or partial hazard observations (e.g., perturbed pocket locations, missed detections, false positives, measurement noise) and report how the planner's advantage degrades.
  2. [Section IV-C, Fig. 5] The central quantitative claim rests on mean costs over 500 trials, but the paper reports only table means with no standard deviations, confidence intervals, or significance tests. The reader cannot determine whether the claimed improvements of 16.3%, 26.7%, 26.2% (and the larger percentages for 24/36 PoIs) are statistically reliable. Please provide per-condition variance estimates (e.g., standard errors or bootstrap confidence intervals) and a paired significance test between the proposed planner and each baseline.
  3. [Section III-A.2, Eq. (4)] The POMDP-to-MDP reduction assumes that the likelihoods P(l) are fixed and that inspecting one PoI does not update beliefs about other PoIs. Under the generative model of Eq. (5), damage is driven by shared high-wind pockets, so when pocket locations are uncertain, observing damage at one PoI is informative about other PoIs. The experiments avoid this issue by giving the GNN exact wind-pocket node locations, which makes the fixed-P approximation near ground truth. The authors should either justify the fixed-P assumption in the partially observed setting or test a variant with unknown/uncertain pocket locations and belief updating, and show that the planner still performs well.
  4. [Section IV-D] The real-world demonstration is a single scenario compared only against the optimistic baseline, and the reported 20.43% improvement has no repetitions or error bars. This is acceptable as a feasibility demonstration, but it cannot serve as evidence for the headline improvement claims. Please either label this explicitly as a feasibility study or provide additional real-world trials.
minor comments (4)
  1. [Section IV-C.2] The sentence 'our model outperforms the baselines by 34.0%, 39.9%, and 37.3%' is not tied to a specific environment size; clarify whether these numbers refer to 24 PoIs, 36 PoIs, or an aggregate, since Fig. 5 reports six separate conditions.
  2. [Section IV-C] The learned greedy baseline is underspecified for multi-robot settings: the text says the team is 'assigned to the PoI with high likelihood' but does not explain how multiple robots are coordinated or whether travel cost enters the selection. Please provide the exact assignment rule used.
  3. [Section III-A.2, Eq. (4)] The state transition after the first PoI is inspected is not fully defined: the paper states that the robots move to 'new locations qt+1' but does not specify how the positions of robots that have not finished their actions are computed. Please define the transition, or state explicitly that replanning occurs at each reveal event and qt+1 is the measured current pose.
  4. [Section IV-B] Please clarify the train/test split: 27,000 environments are used for training, and 500 'previously unseen' environments are used for evaluation; state explicitly that the test environments are disjoint from the training set and describe how the procedural generator's randomness is controlled.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the GNN is trained and evaluated on disjoint procedurally generated environments, the planner's objective and evaluation metric are aligned rather than tautological, and the one self-citation is not load-bearing.

full rationale

The claimed derivation chain is not circular. Damage labels are generated from Eq. (5) using random wind-pocket locations; a GNN is trained on 27,000 procedurally generated environments and evaluated on 500 unseen environments; the planner then minimizes the expected-cost recursion Eq. (4) with the GNN's probability estimates; and the reported costs are accumulated from independently sampled Bernoulli damage outcomes, not from the planner's own estimates. The optimization objective (Eq. 4) and the evaluation metric (Eq. 1) are aligned, which is appropriate for a decision-making evaluation rather than a tautology. The only self-citation, [23], supplies the temporally-extended-action abstraction and is non-load-bearing: Eqs. (3)-(4) fully define the transition and cost model in this paper, and no result from [23] is invoked as a proof or uniqueness theorem. A separate concern is that the simulated GNN input includes nodes for the true wind-gust pockets and distance edge features, so the 'noisy sensor data' pathway described in the abstract is not directly exercised; this is an external-validity limitation, not a circular reduction, because the planner's improvement is still contingent on the GNN's learned outputs on held-out environments.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests mainly on a synthetic generative scenario and a modeling simplification. The method itself introduces no new physical or virtual entities; the hand-chosen horizon heuristic and the MDP assumption are the main things the reader pays for upstream.

free parameters (3)
  • Receding-horizon subset size N_PoI = 12
    Hand-selected horizon size used in all experiments; no sensitivity analysis is reported, so the reported gains may depend on it.
  • L_priority composition = 6 most likely + 6 nearest
    Hand-designed heuristic for selecting which PoIs the planner optimizes over; not compared with alternative selection rules.
  • GNN training hyperparameters = 10 epochs, lr 5e-4, hidden dim 16
    Standard choices needed for reproduction, but not fitted to the evaluation metric.
assumptions (6)
  • standard math Standard Bellman optimality and dynamic programming for the MDP in Eq. (2).
    Unproved background used to write Eq. (4).
  • domain assumption Damage likelihood for each PoI follows the Gaussian function in Eq. (5) with sigma=60 and fixed susceptibility values.
    Used to generate both training and evaluation environments; the reported performance is measured within this synthetic distribution.
  • domain assumption PoI damage outcomes are independent Bernoulli samples given the wind-pocket configuration.
    Implied by the sampling procedure; this is what makes fixed P(l) valid without belief updates in the simulator.
  • ad hoc to paper The partially observable search problem can be represented as an MDP with fixed expected costs (Sec. III-A.2).
    This is the paper's key modeling move: it assumes no belief update after inspections and that temporally extended actions terminate when any robot finishes. Proven optimal only under these restrictions, not for general POMDPs.
  • domain assumption The GNN's estimated probabilities are accurate enough to substitute for true P(l) in the planning objective.
    Planner optimality in Eq. (4) depends on P(l) being the true damage probabilities; in simulation the GNN must recover Eq. (5) from node and edge features.
  • domain assumption Inspection time T_R(l) is constant at 30 for all PoIs.
    Simplifies the simulator; if inspection times vary by location, planner behavior and the reported gains could differ.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning-Augmented Model-Based Multi-Robot Planning for Time-Critical Search and Inspection Under Uncertainty." pith.science (2026). https://pith.science/paper/HGRDSNIM

@misc{pith2026250706129,
  author       = {Pith},
  title        = {Pith review of: Learning-Augmented Model-Based Multi-Robot Planning for Time-Critical Search and Inspection Under Uncertainty},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HGRDSNIM}},
  note         = {Machine review of arXiv:2507.06129}
}
read the original abstract

In disaster response or surveillance operations, quickly identifying areas needing urgent attention is critical, but deploying response teams to every location is inefficient or often impossible. Effective performance in this domain requires coordinating a multi-robot inspection team to prioritize inspecting locations more likely to need immediate response, while also minimizing travel time. This is particularly challenging because robots must directly observe the locations to determine which ones require additional attention. This work introduces a multi-robot planning framework for coordinated time-critical multi-robot search under uncertainty. Our approach uses a graph neural network to estimate the likelihood of PoIs needing attention from noisy sensor data and then uses those predictions to guide a multi-robot model-based planner to determine the cost-effective plan. Simulated experiments demonstrate that our planner improves performance at least by 16.3\%, 26.7\%, and 26.2\% for 1, 3, and 5 robots, respectively, compared to non-learned and learned baselines. We also validate our approach on real-world platforms using quad-copters.

Figures

Figures reproduced from arXiv: 2507.06129 by the authors.

Figure 1
Figure 1. An overview of our scenario and approach Learning informs multi-robot model-based planning for effecting planning for time-critical search and inspection under uncertainty. An optimistic planner might prioritize the closest PoIs, lead￾ing to increased time to discover the locations that require immediate response. Similarly, route optimization techniques like the Traveling Salesman Problem (TSP) [1] and its multi￾ro… view at source ↗
Figure 2
Figure 2. Schematic of our Multi-robot Planning Framework A schematic showing how the cost of joint-action under a policy is calculated using our planning framework. For action at, the robot team concurrently explores multiple PoIs until one of the robots finishes inspecting a PoI, leading to a transition in the state. Under our formulation, the state is defined as a tuple st = {L, qt}, where L is the set of un-inspected PoIs… view at source ↗
Figure 3
Figure 3. Example trials We show the trajectory for optimistic, learned greedy, and our approach in a simulated damage survey scenario. The bottom plot shows distance vs net cost and remaining number of damaged locations for each planner. pocket s and the point of interest l, σ = 60 is the scale factor governing the influence range of the high-wind, and Psusceptible is inherent vulnerability to damage of the PoI with values o… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results: 12 and 36 PoIs We compare the cost of our multi￾robot model-based learned planner vs learned and non-learned baselines for 500 experiments in the scatterplot. The statistics for all the results show the benefit of our learning-augmented model-based planner. an…
Figure 5
Figure 5. Figure 5: Results for 12, 24 and 36 PoIs The table shows the average cost accrued in 500 experiments for each planner. The plot net cost and number of remaining damaged locations vs. distance to visit all PoIs covered for our planner (green), learned-greedy baseline (orange), an…
Figure 6
Figure 6. Figure 6: Demonstration of feasibility in deployment of our planner in two quadcopters The figure shows our quadcopter platforms (top left). Bottom left shows the environment where the experiment is done, where PoIs are boxes, circled in yellow. These PoIs denote locations that …

Discussion (0). Continue with ORCID 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

  1. DRIVE: Dynamic Rule Inference and Verified Evaluation for Constraint-Aware Autonomous Driving

    cs.RO 2025-08 unverdicted novelty 5.0 of 10

    DRIVE uses exponential-family likelihoods to learn soft driving constraints from expert data and injects them into convex optimization, reporting 0.0% constraint violations on inD, highD, and RoundD.

Reference graph

Works this paper leans on

26 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    Traveling salesman problem,

    K. L. Hoffman, M. Padberg, G. Rinaldi et al. , “Traveling salesman problem,” Encyclopedia of operations research and management sci- ence, vol. 1, pp. 1573–1578, 2013

    work page 2013

  2. [2]

    Move and im- prove: A distributed multi-robot coordination approach for multiple depots multiple travelling salesmen problem,

    O. Cheikhrouhou, A. Koub ˆaa, and H. Bennaceur, “Move and im- prove: A distributed multi-robot coordination approach for multiple depots multiple travelling salesmen problem,” in 2014 IEEE Interna- tional Conference on Autonomous Robot Systems and Competitions (ICARSC). IEEE, 2014, pp. 28–35

    work page 2014

  3. [3]

    An analytical hierarchy process-based approach to solve the multi- objective multiple traveling salesman problem,

    S. Trigui, O. Cheikhrouhou, A. Koubaa, A. Zarrad, and H. Youssef, “An analytical hierarchy process-based approach to solve the multi- objective multiple traveling salesman problem,” Intelligent Service Robotics, vol. 11, no. 4, pp. 355–369, 2018

    work page 2018

  4. [4]

    Where will the next emergency event occur? predicting ambulance demand in emergency medical services using artificial intelligence,

    G. Grekousis and Y . Liu, “Where will the next emergency event occur? predicting ambulance demand in emergency medical services using artificial intelligence,” Computers, Environment and Urban Systems , vol. 76, pp. 110–122, 2019

    work page 2019

  5. [5]

    Multi-robot path planning for mobile sensing through deep reinforcement learning,

    Y . Wei and R. Zheng, “Multi-robot path planning for mobile sensing through deep reinforcement learning,” in IEEE INFOCOM 2021-IEEE Conference on Computer Communications . IEEE, 2021, pp. 1–10

    work page 2021

  6. [6]

    Multi-robot infor- mation gathering for precision agriculture: Current state, scope, and challenges,

    A. Dutta, S. Roy, O. P. Kreidl, and L. B ¨ol¨oni, “Multi-robot infor- mation gathering for precision agriculture: Current state, scope, and challenges,” IEEE Access, vol. 9, pp. 161 416–161 430, 2021

    work page 2021

  7. [7]

    Multiagent reinforcement learning for multi-robot systems: A survey,

    D. Gu and E. Yang, “Multiagent reinforcement learning for multi-robot systems: A survey,” Technical Report of the Department of Computer Science, 2004

    work page 2004

  8. [8]

    Reinforcement learning in robotics: A survey,

    J. Kober, J. A. Bagnell, and J. Peters, “Reinforcement learning in robotics: A survey,” The International Journal of Robotics Research , vol. 32, no. 11, pp. 1238–1274, 2013

    2013

Show all 26 references
  1. [9]

    Multi-robot task allocation in disaster response: Addressing dynamic tasks with deadlines and robots with range and payload constraints,

    P. Ghassemi and S. Chowdhury, “Multi-robot task allocation in disaster response: Addressing dynamic tasks with deadlines and robots with range and payload constraints,” Robotics and Autonomous Systems , vol. 147, p. 103905, 2022

    2022

  2. [10]

    Multiple uav cooperative searching operation using polygon area decomposition and efficient coverage algorithms,

    I. Maza and A. Ollero, “Multiple uav cooperative searching operation using polygon area decomposition and efficient coverage algorithms,” in Distributed Autonomous Robotic Systems 6 . Springer, 2007, pp. 221–230

    2007

  3. [11]

    Multi-objective multi-robot surveillance,

    F. M. Delle Fave, S. Canu, L. Iocchi, D. Nardi, and V . A. Ziparo, “Multi-objective multi-robot surveillance,” in 2009 4th International Conference on Autonomous Robots and Agents . IEEE, 2009, pp. 68–73

    2009

  4. [12]

    A review on multirobot systems in agriculture,

    C. Ju, J. Kim, J. Seol, and H. I. Son, “A review on multirobot systems in agriculture,” Computers and Electronics in Agriculture, vol. 202, p. 107336, 2022

    2022

  5. [13]

    Comparison of task-allocation algorithms in frontier-based multi-robot exploration,

    J. Faigl, O. Simonin, and F. Charpillet, “Comparison of task-allocation algorithms in frontier-based multi-robot exploration,” in Multi-Agent Systems: 12th European Conference, EUMAS 2014, Prague, Czech Re- public, December 18-19, 2014, Revised Selected Papers 12. Springer, 20...

    2014

  6. [14]

    A novel method for multiple depot and open paths, multiple traveling salesmen problem,

    X. Wang, D. Liu, and M. Hou, “A novel method for multiple depot and open paths, multiple traveling salesmen problem,” in 2013 IEEE 11th international symposium on applied machine intelligence and informatics (SAMI). IEEE, 2013, pp. 187–192

    2013

  7. [15]

    Frontier-based exploration using multiple robots,

    B. Yamauchi, “Frontier-based exploration using multiple robots,” in Proceedings of the second international conference on Autonomous agents, 1998, pp. 47–53

    1998

  8. [16]

    Topological con- straints in search-based robot path planning,

    S. Bhattacharya, M. Likhachev, and V . Kumar, “Topological con- straints in search-based robot path planning,” Autonomous Robots , vol. 33, pp. 273–290, 2012

    2012

  9. [17]

    Multi-robot human guidance using topological graphs,

    P. Khandelwal and P. Stone, “Multi-robot human guidance using topological graphs,” in AAAI Spring 2014 Symposium on Qualitative Representations for Robots (AAAI-SSS) , 2014

    2014

  10. [18]

    Persistent homology for path planning in uncertain environments,

    S. Bhattacharya, R. Ghrist, and V . Kumar, “Persistent homology for path planning in uncertain environments,” IEEE Transactions on Robotics, vol. 31, no. 3, pp. 578–590, 2015

    2015

  11. [19]

    Scalable and exact milp methods for uav persistent visitation prob- lem,

    K. Kalyanam, S. Manyam, A. V on Moll, D. Casbeer, and M. Pachter, “Scalable and exact milp methods for uav persistent visitation prob- lem,” in 2018 IEEE Conference on Control Technology and Applica- tions (CCTA). IEEE, 2018, pp. 337–342

    2018

  12. [20]

    V oronoi- based multi-robot autonomous exploration in unknown environments via deep reinforcement learning,

    J. Hu, H. Niu, J. Carrasco, B. Lennox, and F. Arvin, “V oronoi- based multi-robot autonomous exploration in unknown environments via deep reinforcement learning,” IEEE Transactions on Vehicular Technology, vol. 69, no. 12, pp. 14 413–14 423, 2020

    2020

  13. [21]

    Allocation using a heterogeneous space voronoi diagram,

    X. Feng and A. T. Murray, “Allocation using a heterogeneous space voronoi diagram,” Journal of Geographical Systems, vol. 20, pp. 207– 226, 2018

    2018

  14. [22]

    Quantitative evaluation of voronoi graph search algorithm in uav path planning,

    C. Zhang, H. Liu, and Y . Tang, “Quantitative evaluation of voronoi graph search algorithm in uav path planning,” in 2018 IEEE 9th International Conference on Software Engineering and Service Science (ICSESS), 2018, pp. 563–567

    2018

  15. [23]

    Learning augmented, multi-robot long- horizon navigation in partially mapped environments,

    A. Khanal and G. J. Stein, “Learning augmented, multi-robot long- horizon navigation in partially mapped environments,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) , 2023, pp. 10 167–10 173

    2023

  16. [24]

    PyBind11 – Seamless operability between C++11 and Python,

    W. Jakob, J. Rhinelander, D. Moldovan and others, “PyBind11 – Seamless operability between C++11 and Python,” 2017

    2017

  17. [25]

    Masked label prediction: Unified message passing model for semi- supervised classification,

    Y . Shi, Z. Huang, S. Feng, H. Zhong, W. Wang, and Y . Sun, “Masked label prediction: Unified message passing model for semi- supervised classification,” in Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21 , Z.-H. Zhou, Ed. Inter...

    2021

  18. [26]

    Self-triggered optimal servicing in dynamic environments with acyclic structure,

    C. Nowzari and J. Cort ´es, “Self-triggered optimal servicing in dynamic environments with acyclic structure,” IEEE Transactions on Automatic Control, vol. 58, no. 5, pp. 1236–1249, 2012

    2012

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.