Pith. sign in

REVIEW 5 major objections 6 minor 21 references

Intercepting an Agile Target with Net-Carrying Drones using Competitive Multi-Agent Reinforcement Learning

T0 review · 5 major / 6 minor · reviewed 2026-07-08 · grok-4.5

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

Pith's one-line read Net-carrying drones learn to intercept an agile flyer by training both sides as competing teams

desk verdict Solid applied MARL for multi-drone net interception under CTBR; the contribution is the combination and ablations, but every performance number hangs on baseline strength and the sim catch model. read the letter →

arxiv 2607.05939 v1 pith:A3BQ3NXS submitted 2026-07-07 cs.RO

Timothée Gavin (ENAC-LAB , LAAS-RIS) , Murat Bronz (ENAC) This is my paper · ORCID
classification cs.RO
keywords multi-agentreinforcementlearningpursuit-evasiondroneinterceptionfictitiousself-playMAPPOCTBRcontrolnet-carryingdronesagileflight
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 shows that a small team of agile drones carrying catching nets can intercept a fast, maneuvering target drone when both sides are trained end-to-end with competitive multi-agent reinforcement learning. Pursuers and the evader are trained together with Multi-Agent Proximal Policy Optimization and Prioritized Fictitious Self-Play, so that each side continually faces a diverse set of past opponent strategies rather than overfitting to the current one. Training runs in a high-fidelity simulator under low-level collective-thrust and body-rate commands, which lets the agents discover agile flight and cooperative tactics instead of relying on high-level velocity setpoints. Against several hand-designed baselines the learned pursuers catch the target more often, catch it faster, and crash less often; ablations show that the self-play curriculum and the low-level control interface are both necessary for those gains. A sympathetic reader cares because the same competitive loop produces emergent cooperation among pursuers and policies that remain effective against opponents they never saw during the final training stage, pointing to a practical route for multi-drone interception without hand-crafted pursuit laws.

What carries the argument

Prioritized Fictitious Self-Play (PFSP) inside MAPPO: each side is trained against a prioritized sampling of past opponent checkpoints so that non-stationarity and catastrophic forgetting of earlier strategies are reduced while both teams keep improving.

What would settle it

Deploy the learned CTBR policies on real net-carrying quadrotors against a human- or autopilot-piloted agile target and measure whether catch rate, time-to-catch, and crash rate still exceed the same heuristic baselines under comparable sensing and actuation limits.

Watch

Extended reading notes

Core claim

Pursuer and evader policies trained jointly with MAPPO and Prioritized Fictitious Self-Play under low-level CTBR control achieve higher catch rates, shorter time-to-catch, and lower crash rates than heuristic baselines in a high-fidelity simulator; PFSP yields policies that remain effective across varied opponent strategies, and cooperative tactics emerge among the pursuers.

Load-bearing premise

The high-fidelity simulator—its dynamics, sensing model, net-catch geometry, and reward design—plus the chosen heuristic baselines form a fair enough testbed that the reported catch, time, and crash gains reflect genuine strategy quality rather than simulator-specific artifacts or weak baselines.

Editorial extensions

If this is right

  • A small fleet of net-carrying drones can intercept an agile single flyer without hand-tuned pursuit laws.
  • Training both pursuers and evader together produces more robust interception policies than training against fixed opponents.
  • Low-level CTBR control is necessary for the agents to discover the agile maneuvers that high-level velocity commands cannot express.
  • Cooperative spatial tactics among pursuers appear spontaneously from the competitive reward, reducing the need to script formation behaviors.

Reading between the lines

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

  • The same PFSP loop could be reused for other competitive multi-robot tasks (dogfighting, ball games, perimeter defense) where both sides must stay agile.
  • If real-world transfer succeeds, the approach offers a path to autonomous counter-UAS systems that adapt to new evasion styles by continuing self-play offline.
  • The emergence of cooperation suggests that simply increasing the number of pursuers under the same reward may yield more sophisticated team plays without extra multi-agent machinery.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The manuscript formulates multi-drone interception of an agile aerial target with net-carrying pursuers as a competitive multi-agent RL problem. Pursuers and a single evader are trained with MAPPO under Prioritized Fictitious Self-Play (PFSP) using low-level collective-thrust and body-rate (CTBR) commands in a high-fidelity simulator. The authors report higher catch rates, lower time-to-catch, and lower crash rates than heuristic baselines; ablations attribute robustness gains to PFSP and performance gains to CTBR versus higher-level commands; and a qualitative analysis claims emergent multi-pursuer cooperative tactics.

Significance. If the reported gains and robustness claims hold under fair baselines and a physically grounded catch/crash model, the work would be a useful empirical contribution to competitive aerial MARL: it couples low-level agile control with multi-agent interception under nonstationarity, and PFSP is a standard, appropriate tool for opponent diversity. Emergent cooperation among net-carrying pursuers would be of practical interest for multi-UAV capture systems. The contribution is primarily empirical and system-level rather than algorithmic novelty; its value therefore hinges on the credibility of the evaluation testbed and the transparency of reward, catch, and baseline definitions.

major comments (5)
  1. The central performance claim (outperformance on catch rate, time-to-catch, and crash rate) is only as strong as the heuristic baselines. The abstract and evaluation narrative assert superiority without establishing that the baselines are competitive for the same net-carrying, multi-agent, CTBR-level task (e.g., whether they use net geometry, formation, crash costs, and comparable sensing/actuation). If baselines are pure-pursuit or single-agent rules that ignore net geometry and multi-agent coordination while RL agents receive shaped multi-agent rewards and full state, the quantitative gaps are not informative. Please define each baseline algorithmically, state their observation/action spaces relative to the RL agents, and report absolute metrics (with variance over seeds) so readers can judge baseline strength.
  2. The net-catch success predicate and crash model are load-bearing for every reported catch/crash number and for the qualitative claim of emergent cooperation. The manuscript must specify the geometric/contact criterion for a successful catch (net deployment dynamics, closing geometry, contact volume, false-positive conditions) and the crash/collision model (inter-agent and agent-environment). An overly permissive intersection test or under-penalized collision model can inflate catch rates and manufacture cooperative-looking behaviors that do not transfer. Provide the exact catch/crash definitions used in the simulator and, if possible, a sensitivity check under tighter catch geometry.
  3. Reward design is a free-parameter surface that directly shapes catch, time, crash, and cooperation outcomes. The paper should state the full reward equation (catch, crash, time, proximity/cooperation terms and their weights), whether rewards are shared or individual for pursuers, and how the evader reward is defined. Without this, reported cooperation and crash reductions cannot be separated from reward shaping. Include the reward specification in the main text or a clearly referenced appendix and discuss sensitivity to major weight choices.
  4. PFSP is claimed to yield more robust policies across opponent strategies, but the evaluation of 'robustness' must be made precise: which held-out opponent policies or strategy classes are used, how the opponent pool and prioritization schedule are constructed, and whether robustness is measured against policies outside the training pool (including the heuristic baselines and frozen past checkpoints). Report the PFSP sampling/prioritization rule and a table of catch/time/crash against each evaluation opponent class so the ablation supports the robustness claim rather than only in-distribution self-play improvement.
  5. Results appear to rest on a single high-fidelity simulator with no external validation of dynamics, sensing, or net physics. For a systems paper this is acceptable if limitations are explicit, but the manuscript should (i) name the simulator and key dynamics/sensor assumptions, (ii) report number of random seeds, means and spreads for all metrics, and any statistical tests, and (iii) discuss failure modes (e.g., when the evader escapes, when pursuers collide, when nets miss under high relative speed). Without seed-level statistics and failure analysis, the outperformance and emergence claims remain under-supported.
minor comments (6)
  1. Define all acronyms on first use in the main text (MAPPO, PFSP, CTBR) and keep notation for pursuer/evader teams consistent across sections and figures.
  2. Add a clear problem-setup figure showing team sizes, net geometry, observation modalities, and action space (CTBR) so the task is self-contained without reading the full methods.
  3. When claiming 'emergent cooperative tactics,' illustrate with trajectory snapshots or rollouts labeled by tactic type, and distinguish cooperation induced by shared reward from cooperation that appears under individual rewards if both are studied.
  4. Report training compute (environment steps, wall-clock, hardware) and network architectures for reproducibility of the MAPPO+PFSP setup.
  5. Clarify whether the evader is trained jointly throughout or against a fixed/pool of pursuer policies, and how symmetry or asymmetry of information is handled.
  6. Ensure all figures reporting catch/time/crash rates include error bars or distributions over seeds and label the exact evaluation protocol in captions.

Simulated Author's Rebuttal

5 responses · 0 unresolved

We thank the referee for a careful and constructive review. The comments correctly identify that the paper’s value is empirical and system-level, and that this value depends on transparent definitions of baselines, catch/crash predicates, rewards, PFSP evaluation, and seed-level statistics. We agree that several of these elements are under-specified in the current draft. We will revise the manuscript to supply exact algorithmic definitions, absolute metrics with variance, the full reward equations, the PFSP prioritization rule and opponent-class results, simulator assumptions, seed counts, and a failure-mode discussion. We do not claim algorithmic novelty beyond the system integration; we believe the requested clarifications will make the empirical claims fully auditable without changing the core findings.

read point-by-point responses

  1. Referee: The central performance claim is only as strong as the heuristic baselines. Please define each baseline algorithmically, state their observation/action spaces relative to the RL agents, and report absolute metrics (with variance over seeds) so readers can judge baseline strength.

    Authors: We agree that baseline strength and fairness must be made fully explicit; the current draft is insufficient on this point. In revision we will (i) give algorithmic pseudocode for every baseline (pure pursuit / intercept-point guidance, formation-constrained pursuit, and any single-agent RL or scripted net-deployment rules we use), (ii) state observation and action spaces side-by-side with the MAPPO agents (including whether baselines receive net geometry, teammate state, crash costs, and CTBR vs. higher-level commands), and (iii) report absolute catch rate, time-to-catch, and crash rate with mean ± std over random seeds rather than only relative gaps. Where a baseline cannot use CTBR or multi-agent net geometry by design, we will say so and justify why it remains a meaningful reference for the literature. We will not claim superiority over baselines that were never given comparable sensing or actuation. These additions will appear in the evaluation section and an appendix table. revision: yes

  2. Referee: The net-catch success predicate and crash model are load-bearing. Specify the geometric/contact criterion for a successful catch (net deployment dynamics, closing geometry, contact volume, false-positive conditions) and the crash/collision model. Provide exact definitions and, if possible, a sensitivity check under tighter catch geometry.

    Authors: The referee is correct that every catch/crash number and the cooperation narrative rest on these predicates, and the manuscript currently under-specifies them. We will add an explicit subsection defining: (a) net geometry and deployment (opening time/shape if any, or fixed aperture), (b) the exact geometric test used for a successful catch (e.g., evader center or body volume intersecting the net plane/volume within a tolerance, relative-speed or approach-angle constraints if any), (c) false-positive guards, and (d) inter-agent and agent–environment collision models (sphere/capsule radii, whether soft contacts or hard terminations, and how crashes are scored). We will also run and report a sensitivity check under a tighter catch volume (and, if feasible, a stricter relative-speed condition) so readers can see how catch rate and emergent tactics degrade. If any cooperative-looking behavior disappears under the tighter test, we will state that plainly. These definitions will be placed in the main text with parameters tabulated. revision: yes

  3. Referee: Reward design is a free-parameter surface. State the full reward equation (catch, crash, time, proximity/cooperation terms and weights), whether rewards are shared or individual for pursuers, and how the evader reward is defined. Include the specification in the main text or a clearly referenced appendix and discuss sensitivity to major weight choices.

    Authors: We agree. Without the full reward, catch/time/crash improvements and ‘cooperation’ cannot be separated from shaping. The revision will include the complete pursuer and evader reward equations (catch bonus, crash penalty, time/step cost, any distance-to-evader or inter-pursuer proximity/cooperation terms, and all scalar weights), and will state clearly whether pursuer rewards are shared, individual, or a mix under MAPPO’s centralized critic. We will place the equations in the main method section (or a short appendix with a main-text pointer) and add a brief sensitivity discussion or ablation on the dominant weights (especially catch vs. crash vs. proximity), reporting how catch rate, crash rate, and qualitative tactics change. We will not claim that cooperation is purely emergent if a proximity/cooperation term is present; we will attribute behavior to the reward structure honestly. revision: yes

  4. Referee: PFSP robustness must be made precise: which held-out opponent policies or strategy classes are used, how the opponent pool and prioritization schedule are constructed, and whether robustness is measured against policies outside the training pool. Report the PFSP sampling/prioritization rule and a table of catch/time/crash against each evaluation opponent class.

    Authors: We accept this criticism. The current claim that PFSP yields ‘more robust’ policies is too vague without an explicit evaluation protocol. We will revise to specify: (i) how the historical opponent pool is built and frozen, (ii) the prioritization/sampling rule (e.g., win-rate or score-based PFSP probabilities and any temperature or mix with latest/self), (iii) the training schedule, and (iv) a held-out evaluation set that includes frozen past checkpoints, the heuristic baselines, and any strategy classes not used for prioritization. We will add a table of catch rate, time-to-catch, and crash rate (mean ± std) of the final pursuer policy against each opponent class, and the symmetric table for the final evader where applicable. Robustness will be claimed only where out-of-pool or cross-strategy metrics support it; in-distribution self-play gains will be labeled as such. The PFSP ablation will be rewritten around this table. revision: yes

  5. Referee: Results rest on a single high-fidelity simulator with no external validation. Name the simulator and key dynamics/sensor assumptions; report number of random seeds, means and spreads for all metrics, and any statistical tests; discuss failure modes (evader escapes, pursuer collisions, nets miss under high relative speed).

    Authors: We agree that for a systems/empirical paper these items are mandatory. The revision will (i) name the simulator and state the dynamics, aerodynamics, actuation (CTBR), sensing, and net-physics assumptions and simplifications, (ii) report the number of independent random seeds for training and evaluation, and give mean ± standard deviation (or other spread) for every metric in the main tables, with statistical tests where we assert significant outperformance, and (iii) add a failure-mode analysis: conditions under which the evader escapes, typical pursuer–pursuer or pursuer–environment collisions, and miss cases at high closing speed or unfavorable net orientation. We will explicitly list sim-to-real and net-physics limitations rather than implying external validation we do not have. We believe this does not undermine the contribution if the evaluation is transparent; it is the appropriate standard for a single-simulator study. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical MAPPO+PFSP training paper with experimental comparisons, not a definitional or self-proving derivation.

full rationale

This is an empirical multi-agent RL robotics paper. The claimed results (higher catch rate, lower time-to-catch and crash rate vs heuristic baselines; PFSP robustness; emergent cooperative tactics under CTBR) are experimental outcomes of training and evaluation in a simulator, not first-principles derivations or fitted constants renamed as predictions. Prioritized Fictitious Self-Play intentionally trains against a pool of past opponent policies; that is standard competitive MARL practice to mitigate nonstationarity and overfitting, not a circular proof that the learned policy is optimal by construction. Ablations (PFSP vs plain self-play; CTBR vs higher-level commands) and baseline comparisons are empirical tests, not self-definitional identities. There is no uniqueness theorem imported from the authors, no ansatz smuggled in as a forced mathematical fact, and no renaming of a known closed-form result as a novel derivation. Concerns about baseline strength or simulator catch/crash model fidelity are validity/correctness risks, not circularity of the derivation chain. Score 0 with empty steps is the appropriate honest finding.

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

Abstract-only audit. The claim rests on standard MARL and quadrotor-control assumptions plus many unspecified training and simulation design choices (rewards, net model, opponent pool, hyperparameters) that function as free parameters. No new physical entities are invented; the ‘entities’ are algorithmic (PFSP opponent prioritization, CTBR policies). Exhaustiveness is limited by missing method detail.

free parameters (4)
  • MAPPO / network / optimization hyperparameters
    Learning rates, discount, GAE lambda, clip, batch sizes, network widths, entropy coefficients, etc. are required for the reported policies but not given in the abstract; performance typically depends on them.
  • Reward function weights (catch, crash, time, cooperation terms)
    Catch rate, time-to-catch, and crash trade-offs are shaped by reward design; abstract does not specify the scalarization, so these weights are free parameters of the claimed behavior.
  • PFSP opponent-pool prioritization and sampling schedule
    Prioritized fictitious self-play depends on how past policies are scored and sampled; that schedule is a design choice that drives the robustness claim.
  • Net geometry / catch success model parameters
    Whether a ‘catch’ registers depends on net size, attachment, contact thresholds, and relative pose tolerances in sim—unspecified free parameters of the success metric.
assumptions (4)
  • domain assumption Competitive multi-agent pursuit-evasion with nets is well-posed as a MARL problem with simultaneous learning of pursuers and evader.
    Abstract formulates the task as competitive MARL; this assumes shared observability/action interfaces and a meaningful zero-sum or mixed objective.
  • domain assumption High-fidelity simulator dynamics and sensing are adequate proxies for agile CTBR flight and net interception evaluation.
    All training and reported metrics are sim-based; transfer or physical validity is not established in the abstract.
  • standard math MAPPO is a valid multi-agent policy-gradient method for this partially competitive team structure.
    Uses established MAPPO algorithm as the learning backbone.
  • ad hoc to paper PFSP sufficiently counters nonstationarity and catastrophic forgetting relative to plain self-play for this task.
    Central methodological bet of the paper; supported only by claimed ablations, not by a general theorem in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Intercepting an Agile Target with Net-Carrying Drones using Competitive Multi-Agent Reinforcement Learning." pith.science (2026). https://pith.science/paper/A3BQ3NXS

@misc{pith2026260705939,
  author       = {Pith},
  title        = {Pith review of: Intercepting an Agile Target with Net-Carrying Drones using Competitive Multi-Agent Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A3BQ3NXS}},
  note         = {Machine review of arXiv:2607.05939}
}
read the original abstract

This article presents a solution to intercept an agile drone by a team of agile drone carrying catching nets. We formulate the problem as a competitive Multi-Agent Reinforcement Learning (MARL) task. To address the problem of nonstationarity and catastrophic forgetting of agents overfitting to the current opponent strategy, we train the pursuers and the evader using Multi-Agent Proximal Policy Optimization (MAPPO) with Prioritized Fictitious Self Play (PFSP). We train the agents in a high-fidelity simulator using low-level control commands, collective thrust and body rates (CTBR), to achieve agile flights for both the pursuers and the evader. We compare the performance of the trained policies in terms of catch rate, time to catch and crash rates, against heuristic baselines and show that our solution outperforms them. Ablation studies show that PFSP lead to more robust policies that can adapt to different opponent strategies, and that a low-level control commands are crucial for learning performing strategies in the pursuit-evasion task. Finally, a qualitative analysis of the learned behaviours highlights the emergence of cooperative tactics among the pursuers.

Figures

Figures reproduced from arXiv: 2607.05939 by the authors.

Figure 1
Figure 1. A multi-agent competitive reinforcement learning approach to train [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the training method from the pursuers’ perspective: In each epoch, one actor network per team is trained with MAPPO and its [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. A successful capture: The three pursuers split up and push the evader towards the inner boundaries, blocking its escape routes and anticipating that [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Two additional examples of the pursuers’ coordinated strategies: in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Two successive interception attempts: (a) the evader narrowly avoids [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    Survey on Anti-Drone Systems: Components, Designs, and Challenges,

    S. Park, et al., “Survey on Anti-Drone Systems: Components, Designs, and Challenges,” IEEE Access, vol. 9, pp. 42 635–42 659, 2021

    work page 2021

  2. [2]

    Yanushevsky, Modern Missile Guidance , 2nd ed

    R. Yanushevsky, Modern Missile Guidance , 2nd ed. Boca Raton: CRC Press, Sept. 2018

    work page 2018

  3. [3]

    Search and pursuit-evasion in mobile robotics,

    T. H. Chung, G. A. Hollinger, and V . Isler, “Search and pursuit-evasion in mobile robotics,” Autonomous Robots, vol. 31, no. 4, pp. 299–316, Nov. 2011

    work page 2011

  4. [4]

    Champion-level drone racing using deep rein- forcement learning,

    E. Kaufmann, et al. , “Champion-level drone racing using deep rein- forcement learning,” Nature, vol. 620, no. 7976, pp. 982–987, 2023

    work page 2023

  5. [5]

    Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm

    D. Silver, et al. , “Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm,” Dec. 2017, arXiv:1712.01815 [cs]

    work page Pith review arXiv 2017

  6. [6]

    Dota 2 with Large Scale Deep Reinforcement Learning

    OpenAI, et al. , “Dota 2 with Large Scale Deep Reinforcement Learning,” Dec. 2019, arXiv:1912.06680 [cs]

    work page Pith review arXiv 2019

  7. [7]

    Towards safe mid-air drone interception: Strategies for tracking & capture,

    M. Pliska, et al., “Towards safe mid-air drone interception: Strategies for tracking & capture,” IEEE Robotics and Automation Letters, 2024

    work page 2024

  8. [8]

    Optimal vehicle- target assignment: A swarm of pursuers to intercept maneuvering evaders based on ideal proportional navigation,

    M. M. Asadi, L. G. Gianoli, and D. Saussi ´e, “Optimal vehicle- target assignment: A swarm of pursuers to intercept maneuvering evaders based on ideal proportional navigation,” IEEE Transactions on Aerospace and Electronic Systems , vol. 58, no. 2, pp. 1316–1332, 2021

    work page 2021

Show all 21 references
  1. [9]

    Distributed guidance for interception by using multiple rotary-wing unmanned aerial vehicles,

    B. Zhu, A. H. B. Zaini, and L. Xie, “Distributed guidance for interception by using multiple rotary-wing unmanned aerial vehicles,” IEEE Transactions on Industrial Electronics, vol. 64, no. 7, pp. 5648– 5656, 2017

    2017

  2. [10]

    Cooperative pursuit with voronoi partitions,

    Z. Zhou, et al. , “Cooperative pursuit with voronoi partitions,” Auto- matica, vol. 72, pp. 64–72, 2016

    2016

  3. [11]

    Learning Multi-Pursuit Evasion for Safe Targeted Navigation of Drones,

    J. Xiao and M. Feroskhan, “Learning Multi-Pursuit Evasion for Safe Targeted Navigation of Drones,”IEEE Transactions on Artificial Intel- ligence, vol. 5, no. 12, pp. 6210–6224, Dec. 2024, arXiv:2304.03443 [cs]

    2024 arXiv

  4. [12]

    Game of Drones: Multi-UA V Pursuit-Evasion Game With Online Motion Planning by Deep Reinforcement Learning,

    R. Zhang, et al., “Game of Drones: Multi-UA V Pursuit-Evasion Game With Online Motion Planning by Deep Reinforcement Learning,” IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 10, pp. 7900–7909, Oct. 2023

    2023

  5. [13]

    Dacoop-a: Decentralized adaptive cooperative pursuit via attention,

    Z. Zhang, et al., “Dacoop-a: Decentralized adaptive cooperative pursuit via attention,” IEEE Robotics and Automation Letters , vol. 9, no. 6, pp. 5504–5511, 2023

    2023

  6. [14]

    Online planning for multi-uav pursuit-evasion in unknown environments using deep reinforcement learning,

    J. Chen, et al. , “Online planning for multi-uav pursuit-evasion in unknown environments using deep reinforcement learning,” 2024

    2024

  7. [15]

    Learned controllers for agile quadrotors in pursuit-evasion games,

    A. S. Roncero, et al. , “Learned controllers for agile quadrotors in pursuit-evasion games,” arXiv preprint arXiv:2506.02849 , 2025

    2025 arXiv

  8. [16]

    Agile interception of a flying target using competitive reinforcement learning,

    T. Gavin, S. Lacroix, and M. Bronz, “Agile interception of a flying target using competitive reinforcement learning,” in Conference on Artificial Intelligence for Defense (CAID) , Rennes, France, 2025

    2025

  9. [17]

    Multi-Agent Actor-Critic for Mixed Cooperative- Competitive Environments,

    R. Lowe, et al. , “Multi-Agent Actor-Critic for Mixed Cooperative- Competitive Environments,” in Advances in Neural Information Pro- cessing Systems, vol. 30. Curran Associates, Inc., 2017

    2017

  10. [18]

    Grandmaster level in starcraft ii using multi-agent reinforcement learning,

    O. Vinyals, et al., “Grandmaster level in starcraft ii using multi-agent reinforcement learning,”nature, vol. 575, no. 7782, pp. 350–354, 2019

    2019

  11. [19]

    The Surprising Effectiveness of PPO in Cooperative Multi-Agent Games,

    C. Yu, et al. , “The Surprising Effectiveness of PPO in Cooperative Multi-Agent Games,” Advances in Neural Information Processing Systems, vol. 35, pp. 24 611–24 624, Dec. 2022

    2022

  12. [20]

    Adaptive incremental nonlinear dynamic inversion for attitude control of micro air vehicles,

    E. J. Smeur, Q. Chu, and G. C. De Croon, “Adaptive incremental nonlinear dynamic inversion for attitude control of micro air vehicles,” Journal of Guidance, Control, and Dynamics , vol. 39, no. 3, pp. 450– 461, 2016

    2016

  13. [21]

    JAX: composable transformations of Python+NumPy programs,

    J. Bradbury, et al. , “JAX: composable transformations of Python+NumPy programs,” 2018

    2018

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.