Pith. sign in

REVIEW 4 major objections 5 minor 29 references

COLMAR establishes that cooperative view policy learning over a shared map, trained with parameter-sharing PPO and a reconstruction-aware reward, substantially outperforms heuristic and non-cooperative baselines for multi-agent active 3D re

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 04:54 UTC pith:HAVH5DFC

load-bearing objection A credible multi-agent extension of active 3D reconstruction with a sensible reward design, but the evaluation has a training-budget confound and weak baselines that need addressing before the headline numbers can be trusted. the 4 major comments →

arxiv 2607.13524 v1 pith:HAVH5DFC submitted 2026-07-15 cs.RO

COLMAR: Cooperative View Policy Learning for Multi-Agent Active 3D Reconstruction

Phu Pham , Damon Conover , Aniket Bera This is my paper
classification cs.RO
keywords multi-agent active reconstructionview planningreinforcement learningPPOTSDF3D Gaussian Splattingcooperative explorationcoverage gain
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that multi-agent active 3D reconstruction can be solved as cooperative policy learning: a single map-centric policy, trained on a fused TSDF volume with a dense reward that rewards unique coverage and team-level discovery while penalizing overlap, stagnation, collisions, and crowding, lets each agent choose its own next view independently at deployment. Cooperation does not come from message passing or a centralized planner; it comes from the shared map and the shared policy, whose reward divides credit among overlapping agents so they learn complementary specializations. If this is right, multi-robot scanning teams can coordinate simply by sharing one fused map and a common policy, without per-agent negotiation or communication bandwidth for decisions. On the Replica benchmark the trained policy reaches 89.4% reconstruction accuracy and 82.6% coverage versus 77.6% and 74.8% for single-agent PPO, with lower Chamfer distance (4.57 cm vs 6.80 cm), and the gains hold across team sizes and sensing budgets. The authors are careful to note the method assumes static scenes and consistent TSDF fusion; sensor noise, calibration errors, and dynamics are declared limitations.

Core claim

The central claim is that an overlap-aware unique-coverage reward — computed as per-step TSDF coverage gain divided by the number of agents concurrently attending the same region, plus a team-level discovery bonus normalized by the number of alive agents — induces a shared PPO policy to allocate complementary viewpoints among agents. The policy, trained with parameter-sharing PPO on a fused TSDF map, executes independently per agent at test time with no inter-agent messaging; coordination emerges from the shared map and shared weights. The authors report that this formulation beats random, greedy coverage, frontier-based, and single-agent PPO baselines under matched sensing budgets and actio

What carries the argument

The shared TSDF volume is the central object: during training it supplies the map-centric observation (rendered depth, ego-centric occupancy, global frontier/free/occupied summary, pose history) and the dense reward; during deployment agents continue to observe the same map features from fused team measurements. The load-bearing reward identity is the overlap-aware unique coverage gain g_ucov = g_lcov / overlap + b_new, where g_lcov is step-wise TSDF coverage gain, overlap measures concurrent multi-agent attention, and b_new encourages first-time discovery; this ratio is the soft credit-allocation mechanism that makes redundant co-observation cheaper and specialization more rewarding. Around

Load-bearing premise

The whole method rests on the assumption that a shared TSDF volume free of sensor noise and calibration errors provides a faithful proxy for reconstruction quality through rendered-depth decreases, so that the reward reliably guides learning.

What would settle it

Inject realistic depth noise and pose drift into the TSDF fusion during training and evaluation, or run the policy on a real RGB-D platform; if the gap over the non-cooperative baseline shrinks or disappears, the proxy assumption is the load-bearing factor. Alternatively, compare against a centralized planner with perfect global coordination; if it matches COLMAR without learning, the learned cooperative policy may be unnecessary.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Under matched sensing budgets, cooperative map-centric training yields higher accuracy and coverage and lower Chamfer than heuristic and single-agent baselines on Replica, with up to 54% accuracy and 49% coverage gains reported.
  • Performance scales with team size: coverage rises from 72.4% (N=1) to 80.9% (N=4) on GLEAM, with Chamfer dropping from 6.48 cm to 4.96 cm, though gains saturate.
  • Longer sensing horizons monotonically improve coverage and Chamfer (from T=50 to T=300), indicating the policy budgets its exploration efficiently.
  • The policy zero-shot transfers from GLEAM to Replica with slightly better Chamfer (4.57 vs 4.96 cm), suggesting learned coordination generalizes across scene types.
  • Ablations confirm that every reward component matters; removing unique coverage and team bonus degrades accuracy and Chamfer, and removing all penalties doubles the collision rate (from 0.07 to 0.14).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: If the rendered-depth proxy for reconstruction quality holds, the same overlap-divided credit scheme could be applied to other multi-agent active perception problems (search, inspection, coverage) where redundant sensing is wasteful.
  • Editorial inference: The paper trains only on clean TSDF inputs; injecting calibrated depth noise or pose drift during training could make the policy robust to the very real-world conditions flagged as limitations, a direct testable extension.
  • Editorial inference: Since deployment needs no decision-time messaging, the coordination bottleneck in bandwidth-limited teams moves to map fusion itself; a sparse or compressed map-sharing protocol would be the next practical requirement.
  • Editorial inference: The saturation beyond four agents suggests the current overlap computation may need to become pairwise or regional to scale to larger teams, another extension the paper does not explore.

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

4 major / 5 minor

Summary. The paper proposes COLMAR, a multi-agent active 3D reconstruction framework in which a shared map-centric policy is trained with parameter-sharing PPO and a dense, reconstruction-aware reward derived from TSDF updates. The reward encourages overlap-aware unique coverage, team-level discovery, and collision-safe exploration. At deployment, agents act independently using the frozen policy and the fused team map; selected trajectories are evaluated with 3D Gaussian Splatting for photometric quality. Experiments on GLEAM and Replica report improvements over random, greedy, frontier, and single-agent PPO baselines, together with ablations of reward components, training regime, and input scope. The authors explicitly acknowledge limitations including static scenes, sensor noise, calibration errors, and domain shift.

Significance. If the reported results hold, COLMAR offers a practical and well-specified recipe for cooperative view planning: shared TSDF-based observations, dense reward shaping from incremental reconstruction updates, and decentralized execution without inter-agent messaging. A notable strength is that final evaluation uses photometric and geometric metrics (PSNR, SSIM, LPIPS, Chamfer) that are not directly optimized by the RL reward, which mitigates the concern that the gains are merely an artifact of optimizing the headline coverage metric. The ablations and cross-dataset transfer are useful. However, the empirical claims are currently supported only by point estimates with no uncertainty quantification, the non-cooperative baseline may be data-starved under the stated training protocol, and the comparison set omits the multi-agent coordination methods discussed in the related work. No code or machine-checked proofs are provided; the contribution is empirical and needs stronger experimental controls to justify the central claim.

major comments (4)
  1. [§IV.C and §VI.B, Tables I and VI] The non-cooperative comparison is confounded with training data volume. Section IV.C specifies 128 rollout steps per update; with N agents, a multi-agent rollout collects 128×N transitions per update while a single-agent rollout collects 128. If the number of updates is the same, the single-agent baseline (Table I) and the single-agent-trained variant (Table VI) see N times less data. The paper does not state that total transitions were matched. This makes it impossible to separate the benefit of joint multi-agent training from a scale effect. Please specify the update/transition budget and, ideally, train single-agent with N× more updates or add a multi-agent-trained decentralized baseline.
  2. [§IV.B and §II.B] The baseline set excludes the multi-agent coordination methods discussed in related work. Only random, greedy, frontier, and single-agent PPO are compared; no non-learning cooperative planner (e.g., the MAC-Ego3D/CORE/collaborative dense mapping systems cited in §II.B) is evaluated. Since the central claim is that cooperative view policy learning improves multi-agent reconstruction, the experiments need at least one strong coordination-aware baseline, or the claim should be explicitly scoped to the compared baselines.
  3. [§IV.D and Tables I–VII] All results are point means without variance. Section IV.D says 'multiple evaluation episodes' but no standard deviations, confidence intervals, or per-scene results are given. With Replica's 18 scenes, differences such as 74.8% vs 77.6% accuracy or 6.80 vs 5.89 cm Chamfer are not shown to be statistically reliable. Add error bars/per-scene breakdowns and, where feasible, significance tests to support the 'consistent improvements' claim.
  4. [§V.E, Fig. 5, and Eq. (9)–(11)] The claim that agents 'spread to complementary frontiers' is supported only qualitatively. No quantitative coordination statistic (e.g., overlap ratio of TSDF updates, spatial entropy, fraction of redundant observations) is reported, even though the reward explicitly penalizes clustering and overlapping coverage. A deployment-time redundancy metric would verify whether the policy actually learns coordination or simply follows the shaped reward. Please report such a metric.
minor comments (5)
  1. [§III.C, Eq. (1)] The equation omits the MLP/concat step described in the text: the text says pose history is encoded by a GRU, concatenated with current pose, and mapped by an MLP, but Eq. (1) writes only h_vec = GRU(x_temp). Please align the equation with the description.
  2. [§II.D, Ref. [24]] The citation [24] appears to be the wrong paper: it is titled 'MAPPO: Maximum A Posteriori Preference Optimization,' not the multi-agent PPO method referenced in the text. Please check and correct the citation.
  3. [§I, Refs. [8]–[10]] The phrase 'MNE-SLAM [8], [9], [10]' should list the systems separately, as [9] is MAGiC-SLAM and [10] is GRAND-SLAM, not MNE-SLAM.
  4. [Fig. 4] The caption says black patches 'do not reflect reconstruction errors.' If these artifacts appear in rendered depth or RGB, please clarify how the photometric and depth metrics treat them.
  5. [References] Reference [17] reports page range '8913–8913'; this looks like a typesetting or indexing error and should be verified.

Circularity Check

0 steps flagged

No significant circularity: COLMAR's claims are grounded by independent geometry and photometric metrics, not by construction.

full rationale

COLMAR's derivation chain is an empirical RL pipeline rather than a derivation from fitted parameters. The policy is trained with PPO on a TSDF-derived reward (Eqs. 9-12), and the paper's headline improvements are measured on metrics that are not the training objective: Chamfer distance against ground truth, Accuracy with tau=5cm, PSNR/SSIM/LPIPS, and depth L1 from 3DGS reconstructions (Section IV.D, Table I). Thus, even though the 'coverage gain' reward is aligned with the reported coverage metric, the photometric and geometric gains are externally grounded and not forced by the reward definition. Self-citations [17,18,22] are confined to related work and are not load-bearing; there is no invoked uniqueness theorem, no ansatz smuggled in via citation, and no fitted parameter renamed as a prediction. The closest concern is the coverage metric being partially aligned with the reward, but the paper also reports independent metrics, so this does not constitute circularity. The single-agent baseline's training-data volume is a potential experimental confound, not a definitional reduction; the paper's own limitations (sensor noise, calibration, domain shift) are honesty statements that further reduce any appearance of circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on a standard RL pipeline plus hand-chosen reward terms. The reward weights and penalty coefficients are not reported; the overlap computation is underspecified; the action discretization and accuracy threshold are design choices. The method assumes static, noise-free scenes where TSDF depth decrease proxies reconstruction quality, and assumes that independent sampling from a shared policy on a shared map yields useful coordination. No new physical entities are introduced.

free parameters (5)
  • Reward weights w_ucov, w_team, b_new = not reported
    Hand-chosen constants in Eqs. (9)-(12) controlling relative strength of unique coverage, team bonus, and new-cell bonus; ablations show they affect performance.
  • Penalty coefficients for p_stag, p_col, p_crowd = not reported
    Hand-chosen scales for stagnation, collision, and crowding penalties in Eq. (9); values omitted.
  • Overlap denominator normalization = implicit overlap_i measure
    Eq. (11) divides local coverage gain by overlap_i; the exact overlap computation/threshold is not specified.
  • Action discretization increments = 0.25 m translation, 15° rotation
    Section III-D: chosen action space; fixed across methods, but a design choice.
  • Accuracy threshold tau = 5 cm
    Section IV-D: evaluation threshold for Accuracy metric; chosen by authors.
axioms (5)
  • domain assumption Static scene assumption
    Section VII: method limited to static indoor scenes; dynamic changes degrade performance.
  • domain assumption Shared, noise-free TSDF fusion with consistent map updates
    Reward and occupancy features rely on shared TSDF with no sensor noise/calibration error; limitations admit sensor noise may degrade.
  • ad hoc to paper Incremental view-space gain (d_after < d_before) is a valid proxy for reconstruction quality
    Eqs. (10)-(11): reward uses this proxy; not proven equivalent to final Chamfer/PSNR.
  • ad hoc to paper Parameter-sharing PPO with independent action selection yields emergent coordination via shared map
    Central assumption of method; no theoretical guarantee; results support empirically.
  • standard math Standard PPO/GAE convergence assumptions
    Uses standard clipped PPO objective Eqs. (5)-(8); assumes standard RL stability conditions.

pith-pipeline@v1.3.0-alltime-deepseek · 9909 in / 11886 out tokens · 110573 ms · 2026-08-02T04:54:40.719345+00:00 · methodology

0 comments
read the original abstract

Active 3D reconstruction requires selecting informative viewpoints under limited sensing budgets. In multi-agent settings, coordination inefficiencies such as redundant observations and spatial clustering can significantly reduce reconstruction quality. We present COLMAR, a cooperative view policy learning framework for multi-agent active 3D reconstruction. COLMAR formulates viewpoint allocation as a shared policy optimization over map-centric observations and introduces a reconstruction-aware objective that promotes overlap-aware coverage, team-level discovery, and collision-safe exploration. Dense feedback derived from incremental reconstruction updates aligns exploration behavior with downstream geometric quality. The policy is trained using parameter-sharing Proximal Policy Optimization (PPO) with independent per-agent action selection at deployment, conditioned on a fused team map and without inter-agent message passing for decision making. Selected viewpoints are then reconstructed with 3D Gaussian Splatting (3DGS) for high-fidelity photometric evaluation. Experiments on GLEAM and Replica demonstrate consistent improvements over heuristic and non-cooperative baselines, achieving up to 54% higher reconstruction accuracy and 49% greater coverage under matched sensing budgets.

Figures

Figures reproduced from arXiv: 2607.13524 by Aniket Bera, Damon Conover, Phu Pham.

Figure 1
Figure 1. Figure 1: Illustration of COLMAR for cooperative multi-agent active 3D [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of COLMAR. During training, a shared TSDF provides map-centric observations and reconstruction-aware rewards for PPO. During [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training dynamics of COLMAR shown as running averages over [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Step-wise qualitative reconstruction progress under COLMAR at Steps 10, 100, and 200. Occupancy maps show explored space (red: occupied, light [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Agent-wise qualitative reconstruction under COLMAR (two snapshots per agent). Columns are grouped by agent to visualize role specialization. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 6 linked inside Pith

  1. [1]

    A frontier-based approach for autonomous exploration,

    B. Yamauchi, “A frontier-based approach for autonomous exploration,” inProceedings 1997 IEEE International Symposium on Computational Intelligence in Robotics and Automation CIRA’97. ’Towards New Computational Principles for Robotics and Automation’, 1997, pp. 146–151

    1997

  2. [2]

    Receding horizon

    A. Bircher, M. Kamel, K. Alexis, H. Oleynikova, and R. Siegwart, “Receding horizon ”next-best-view” planner for 3d exploration,” in 2016 IEEE International Conference on Robotics and Automation (ICRA), 2016, pp. 1462–1468

    2016

  3. [3]

    Gennbv: Generalizable next-best-view policy for active 3d reconstruction,

    X. Chen, Q. Li, T. Wang, T. Xue, and J. Pang, “Gennbv: Generalizable next-best-view policy for active 3d reconstruction,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024, pp. 16 436–16 445

    2024

  4. [4]

    Gleam: Learning generalizable exploration policy for active mapping in com- plex 3d indoor scene,

    X. Chen, T. Wang, Q. Li, T. Huang, J. Pang, and T. Xue, “Gleam: Learning generalizable exploration policy for active mapping in com- plex 3d indoor scene,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2025, pp. 5558– 5568

    2025

  5. [5]

    Nice-slam: Neural implicit scalable encoding for slam,

    Z. Zhu, S. Peng, V . Larsson, W. Xu, H. Bao, Z. Cui, M. R. Oswald, and M. Pollefeys, “Nice-slam: Neural implicit scalable encoding for slam,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022

    2022

  6. [6]

    Nerf-slam: Real-time dense monocular slam with neural radiance fields,

    A. Rosinol, J. J. Leonard, and L. Carlone, “Nerf-slam: Real-time dense monocular slam with neural radiance fields,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 3437–3444

    2023

  7. [7]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, pp. 139:1–139:14, 2023

    2023

  8. [8]

    Mne-slam: Multi-agent neural slam for mobile robots,

    T. Deng, G. Shen, C. Xun, S. Yuan, T. Jin, H. Shen, Y . Wang, J. Wang, H. Wang, D. Wang, and W. Chen, “Mne-slam: Multi-agent neural slam for mobile robots,” in2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 1485–1494

    2025

  9. [9]

    MAGiC-SLAM: Multi-agent gaussian globally consistent slam,

    V . Yugay, T. Gevers, and M. R. Oswald, “MAGiC-SLAM: Multi-agent gaussian globally consistent slam,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 6741–6750

    2025

  10. [10]

    Grand-slam: Local optimization for globally consistent large-scale multi-agent gaussian slam,

    A. Thomas, A. Sonawalla, A. Rose, and J. P. How, “Grand-slam: Local optimization for globally consistent large-scale multi-agent gaussian slam,”IEEE Robotics and Automation Letters, vol. 10, no. 12, pp. 13 129–13 136, 2025

    2025

  11. [11]

    Activegamer: Active gaussian mapping through efficient rendering,

    L. Chen, H. Zhan, K. Chen, X. Xu, Q. Yan, C. Cai, and Y . Xu, “Activegamer: Active gaussian mapping through efficient rendering,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 16 486–16 497

    2025

  12. [12]

    Collaborative large-scale dense 3d reconstruction with online inter-agent pose optimisation,

    S. Golodetz, T. Cavallari, N. A. Lord, V . A. Prisacariu, D. W. Murray, and P. H. S. Torr, “Collaborative large-scale dense 3d reconstruction with online inter-agent pose optimisation,”IEEE Transactions on Visualization and Computer Graphics, vol. 24, no. 11, pp. 2895–2905, 2018

    2018

  13. [13]

    Mac- ego3d: Multi-agent gaussian consensus for real-time collaborative ego- motion and photorealistic 3d reconstruction,

    X. Xu, F. Xue, S. Zhao, Y . Pan, S. Scherer, and X. Huang, “Mac- ego3d: Multi-agent gaussian consensus for real-time collaborative ego- motion and photorealistic 3d reconstruction,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 854– 863

    2025

  14. [14]

    Online next-best-view planner for 3d-exploration and inspection with a mobile manipulator robot,

    M. Naazare, F. G. Rosas, and D. Schulz, “Online next-best-view planner for 3d-exploration and inspection with a mobile manipulator robot,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 3779– 3786, 2022

    2022

  15. [15]

    A shadowcasting-based next-best-view planner for autonomous 3d ex- ploration,

    A. Batinovic, A. Ivanovic, T. Petrovic, and S. Bogdan, “A shadowcasting-based next-best-view planner for autonomous 3d ex- ploration,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2969–2976, 2022

    2022

  16. [16]

    Coordinated multi-robot exploration,

    W. Burgard, M. Moors, C. Stachniss, and F. Schneider, “Coordinated multi-robot exploration,”IEEE Transactions on Robotics, vol. 21, no. 3, pp. 376–386, 2005

    2005

  17. [17]

    Optimizing crowd-aware multi-agent path finding through local communication with graph neural networks,

    P. Pham and A. Bera, “Optimizing crowd-aware multi-agent path finding through local communication with graph neural networks,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 8913–8913

    2024

  18. [18]

    Dream: Decentralized re- inforcement learning for exploration and efficient energy management in multi-robot systems,

    D. Patel, P. Pham, K. Tiwari, and A. Bera, “Dream: Decentralized re- inforcement learning for exploration and efficient energy management in multi-robot systems,”arXiv preprint arXiv:2309.17433, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Core: Coopera- tive reconstruction for multi-agent perception,

    B. Wang, L. Zhang, Z. Wang, Y . Zhao, and T. Zhou, “Core: Coopera- tive reconstruction for multi-agent perception,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 8710–8720

    2023

  20. [20]

    Learning to explore using active neural slam,

    D. S. Chaplot, D. Gandhi, S. Gupta, A. Gupta, and R. Salakhutdinov, “Learning to explore using active neural slam,” inInternational Conference on Learning Representations (ICLR), 2020

    2020

  21. [21]

    Gaussian Splatting SLAM,

    H. Matsuki, R. Murai, P. H. J. Kelly, and A. J. Davison, “Gaussian Splatting SLAM,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  22. [22]

    Flashslam: Accelerated rgb-d slam for real-time 3d scene reconstruction with gaussian splatting,

    P. Pham, D. Conover, and A. Bera, “Flashslam: Accelerated rgb-d slam for real-time 3d scene reconstruction with gaussian splatting,”arXiv preprint arXiv:2412.00682, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Multi-agent actor-critic for mixed cooperative-competitive environ- ments,

    R. Lowe, Y . Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch, “Multi-agent actor-critic for mixed cooperative-competitive environ- ments,”Neural Information Processing Systems (NIPS), 2017

    2017

  24. [24]

    Mappo: Maximum a pos- teriori preference optimization with prior knowledge,

    G. Lan, S. Zhang, T. Wang, Y . Zhang, D. Zhang, X. Wei, X. Pan, H. Zhang, D.-J. Han, and C. G. Brinton, “Mappo: Maximum a pos- teriori preference optimization with prior knowledge,”arXiv preprint arXiv:2507.21183, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    The surprising effectiveness of MAPPO in cooperative, multi-agent games,

    C. Yu, A. Velu, E. Vinitsky, Y . Wang, A. M. Bayen, and Y . Wu, “The surprising effectiveness of MAPPO in cooperative, multi-agent games,”CoRR, vol. abs/2103.01955, 2021. [Online]. Available: https://arxiv.org/abs/2103.01955

    Pith/arXiv arXiv 2021

  26. [26]

    Monotonic value function factorisation for deep multi- agent reinforcement learning,

    T. Rashid, M. Samvelyan, C. S. De Witt, G. Farquhar, J. Foerster, and S. Whiteson, “Monotonic value function factorisation for deep multi- agent reinforcement learning,”J. Mach. Learn. Res., vol. 21, no. 1, Jan. 2020

    2020

  27. [27]

    IMPALA: Scalable distributed deep-rl with im- portance weighted actor-learner architectures,

    L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V . Mnih, T. Ward, Y . Doron, V . Firoiu, T. Harley, I. Dunning, S. Legg, and K. Kavukcuoglu, “IMPALA: Scalable distributed deep-rl with im- portance weighted actor-learner architectures,” inProceedings of the International Conference on Machine Learning (ICML), 2018, pp. 1407–1416

    2018

  28. [28]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  29. [29]

    The replica dataset: A digital replica of indoor spaces,

    J. Straub, T. Whelan, L. Ma, and et al., “The replica dataset: A digital replica of indoor spaces,”arXiv preprint arXiv:1906.05797, 2019

    Pith/arXiv arXiv 1906