Pith. sign in

REVIEW 4 major objections 5 minor 54 references

ReCoDe: Reinforcement Learning-based Dynamic Constraint Design for Multi-Agent Coordination

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read ReCoDe claims that keeping an expert optimization controller intact and learning only additional dynamic constraints can outperform pure MARL and hybrid baselines, delivering 18% better reward with 5% of the samples.

desk verdict A clean and genuinely new hybrid-control formulation with valid theory, but the headline performance claim rests on a peak-selection evaluation that needs a proper final-policy comparison before it can be trusted. read the letter →

arxiv 2507.19151 v2 pith:HVXHLRDO submitted 2025-07-25 cs.RO cs.AIcs.LGcs.MA

Michael Amir , Guang Yang , Zhan Gao , Keisuke Okumura , Heedo Woo , Amanda Prorok This is my paper · ORCID
classification cs.ROcs.AIcs.LGcs.MA
keywords reinforcementlearningmulti-agentcoordinationconstrainedoptimizationdynamicconstraintdesigngraphneuralnetworkscontrolbarrierfunctionsconvexquadraticprogramming
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

ReCoDe is a hybrid multi-agent control framework that keeps a user-designed optimization controller and learns an extra, situation-dependent quadratic constraint on top of it, rather than replacing the controller or learning a policy from scratch. The central claim is that this small structural change is enough to combine the reliability of constrained optimization with the adaptability of multi-agent reinforcement learning: in four navigation and consensus tasks, ReCoDe attains roughly 18% more reward than the next-best baseline and trains using about 5% as many samples as pure MARL, while keeping collision rates near zero during training. The paper also argues theoretically that the learned constraint's radius acts as a 'trust dial' between learned and expert control—tightening it lets agents track any safe feasible trajectory, while loosening it lets the expert objective take over when the learned value estimate is uncertain. If these claims hold, the approach matters because it offers a route to adaptive coordination that preserves the safety and interpretability of classical controllers.

What carries the argument

The central object is the learned quadratic constraint $\|u_i(t)-a_i(t)\|^2 \le b_i(t)+s_0$, appended to the handcrafted convex QCQP solved by each agent at every timestep. The policy outputs the reference action $a_i(t)$ and the uncertainty radius $b_i(t)$; the radius determines how much authority the expert objective keeps over the final control input, so the constraint functions as an adjustable trust dial between learned and expert control. Training uses multi-agent PPO with a graph attention network that aggregates messages from neighboring agents, giving decentralized, permutation-invariant inference; slack variables on learned constraints keep the batched solver from failing when a parameter choice makes the problem infeasible. This construction carries the argument because it shifts the learning problem from 'which action?' to 'which feasible set?', preserving existing safety certificates while giving the learner an action space of dimension $m+1$ instead of the full control space.

What would settle it

After training ReCoDe in the Narrow Corridor scenario, collect a large set of observations and evaluate the trained policy's $(a,b)$: if any reachable observation makes the quadratic program infeasible, or if driving $b$ to zero while the expert's preferred action lies outside the handcrafted feasible set produces a violation, the claimed guarantee fails. A direct real-robot variant is to force two teams into the corridor midpoint deadlock and record whether ReCoDe always completes the swap without a solver failure.

Watch

Extended reading notes

Core claim

The paper's discovery, stated on its own terms, is that reinforcement learning can be injected into a constrained-optimization controller at the level of constraints rather than actions or objectives. ReCoDe trains a decentralized graph-neural-network policy to output, at each timestep, a reference action $a(t)$ and an uncertainty radius $b(t)$, which define the added constraint $\|u(t)-a(t)\|^2 \le b(t)$ for the solver's decision variable $u(t)$. Because this ball sits inside the user's original feasible set, the expert controller's safety constraints remain in force; because the ball can shrink or expand, the agent can switch between precise learned control and delegation to the expert. The paper reports that, across its four tasks, this design beats the handcrafted controller, two hybrid baselines, RVO, and pure end-to-end MARL, and that it transfers to a real-robot corridor position-swap where the expert controller alone deadlocks. The accompanying propositions say that a sufficiently tight ball with a large slack penalty forces the solver to track any safe feasible trajectory within $\varepsilon$, and that enlarging the ball can improve true reward when the learned critic is locally flat while the expert objective is decisive.

Load-bearing premise

The load-bearing premise is that the optimization (1) always has a non-empty, convex feasible set for every observation and parameter vector, with a strictly convex objective and a unique, continuously differentiable solution mapping; if a learned constraint ever makes the problem infeasible, the safety and tracking guarantees no longer hold.

Editorial extensions

If this is right

  • If ReCoDe is right, an imperfect expert controller can be improved without surrendering its safety certificates: the original constraints remain in the optimization, so guarantees from the handcrafted layer stay valid during deployment.
  • The ablation result—learning only constraint parameters converges faster and reaches higher reward than learning the objective or both—implies that expert-designed objectives act as a useful inductive bias, and that moving the objective makes learning harder.
  • ReCoDe's sample efficiency (about 5% of pure MARL's samples) and near-zero training collisions suggest that hybrid constraint-learning is usable in settings where exploratory failures are expensive.
  • Because the changed output is only a reference action and a radius, the controller can be executed with local communication and gives a simple, interpretable signal of when the learned policy trusts itself versus the expert.
  • The real-robot corridor swap indicates the approach transfers from simulation to physical robots in at least this narrow-corridor setting, without retraining.

Reading between the lines

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

  • Not in the paper: the uncertainty radius $b(t)$ could double as an online confidence monitor for human oversight, since a large radius marks states where the learned policy defers to the expert; the paper reports the correlation but does not propose this use.
  • Not in the paper: because the policy is built on a permutation-invariant graph network, ReCoDe may generalize to different numbers of agents without retraining; the architecture makes this plausible but the paper does not test it.
  • Not in the paper: the idea of learning constraints rather than objectives should extend to other convex optimization-based controllers (e.g., manipulation or traffic), but the paper explicitly leaves non-navigation settings unstudied, so that extension is a conjecture.
  • Not in the paper: if the learned constraints ever make the optimization infeasible at deployment, the safety guarantee would need a separate fallback mechanism; the paper uses slack variables during training to avoid solver failure, but does not give a formal deployment-time infeasibility certificate.
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, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes ReCoDe, a decentralized hybrid framework for multi-agent navigation and consensus. ReCoDe keeps a handcrafted optimization-based controller (a convex QCQP) and augments it with a single learned quadratic constraint ||u-a||^2 <= b, where the policy (a GNN trained with MAPPO) outputs the reference action a and an uncertainty radius b. This lets the reinforcement-learning policy shape the feasible action set rather than replacing the controller. The paper contributes two theoretical results: Proposition 3.1 shows that, under a feasibility assumption, the proposed constraint parameterization can track any safe feasible trajectory with arbitrarily small error, and Proposition 3.2 gives a bound showing that enlarging b can improve the true Q-value when the learned critic is locally flat and the expert objective is locally decisive. Empirically, the paper evaluates ReCoDe in four scenarios (Narrow Corridor, Connectivity, Waypoint Navigation, Sensor Coverage) against handcrafted controllers, Online CBF, shielding, RVO, and pure MARL, and reports that ReCoDe attains on average 18% better reward than the next-best method and uses 5% of the samples of end-to-end MARL. It also reports a real-robot corridor-swap demonstration and an ablation comparing learning constraints, learning the objective, and learning both.

Significance. If the empirical claims are established, ReCoDe is a valuable contribution to hybrid optimization-based control and multi-agent reinforcement learning. The idea of learning dynamic constraints rather than replacing the controller or shaping the objective is a natural and potentially practical middle ground, and the theoretical propositions, although conditional, are nontrivial and correctly motivate the design. The real-robot demonstration, the safety-preserving training behavior, and the explicit focus on coordination-driven deadlocks are strengths. However, the current evaluation protocol is non-standard (best-of-training-window selection, no per-seed variance, no code release), and the reported mechanism-correlation effect sizes are negligible, so the central empirical claims are not yet convincingly established. The paper does not ship code or data, which further limits verification.

major comments (4)
  1. [Sec. 4.1, Table 2 and Fig. 2 caption] The evaluation protocol is not appropriate for the central claim of empirical superiority. Table 2 reports 'mean +/- standard deviation over best 6 consecutive training steps' across 75 random starting conditions. Selecting the best training window post hoc is a form of checkpoint selection: it does not describe the performance of any fixed, deployable policy, and the standard deviation is over starting conditions, not over independent training runs, so there is no measure of run-to-run variance. In addition, the headline '18% better reward than the next-best method' is the average of per-scenario relative improvements computed on rewards with different units, with an unknown maximum in Sensor Coverage. To support the paper's claims, please report performance of a fixed policy (e.g., the final policy after a predetermined number of steps) averaged over at least 10 independent training seeds with standard errors, and release code and data for reproducibility.
  2. [Sec. 2, Eq. (1) and Sec. 6] The formal safety guarantee relies on the optimization problem (1) being feasible at every time step. The paper introduces slack variables only during training to avoid batch solver failures, but at deployment the agents solve their local QP without explaining how infeasibility of the learned quadratic constraint combined with the handcrafted constraints is avoided. A learned (a,b) that makes the intersection of the ball {u : ||u-a||^2 <= b} with the safety constraints empty would cause the solver to fail and would void the safety claim. Please either prove that the learned policy's output always preserves feasibility, provide a deployment-time fallback (e.g., projecting the reference action onto the handcrafted feasible set before solving), or empirically demonstrate with statistics that infeasibility does not occur during deployment.
  3. [Sec. 4.1, Figs. 3d-e] The mechanistic claim that ReCoDe learns to tighten the uncertainty radius in congested situations and relax it when the path is clear is not supported by the reported correlations. The paper states r approx -0.03 for b vs. number of neighbors and r approx 0.08 for b vs. y-position, with extremely small p-values. These are negligible effect sizes; with large sample sizes, p-values of this magnitude are uninformative. The conclusion 'ReCoDe tightens b to resolve likely deadlocks and relaxes it once the path is clear' does not follow from correlations that are essentially zero. Please provide a more meaningful analysis, such as state-conditioned distributions of b (e.g., b given congestion level), bootstrapped confidence intervals for the effect, or nonparametric tests that quantify effect sizes.
  4. [Sec. 4.1 and Fig. 3b] The claim that ReCoDe 'trains using just 5% as many samples as end-to-end MARL' is based on a single scenario (Waypoint Navigation with agent radius 0.1). The abstract and introduction present this as a general advantage of ReCoDe. Please qualify the claim to the specific scenario where it is measured, or provide a sample-complexity comparison across all four scenarios. Also clarify what '5%' means: steps to reach a given reward threshold, total training budget, or something else.
minor comments (5)
  1. [Throughout] There are typographical errors, including 'Propostion' (Sec. 1), 'adaptible' (Related Work), 'and and' (Sec. 4), and 'over-sized' (Waypoint Navigation).
  2. [Appendix C] The statement that 'with fewer than m+1 linear constraints, it is impossible to specify a bounded region in R^m that contains an epsilon-ball around a point' is imprecise: a single halfspace contains an epsilon-ball, and boundedness is not required for containing a ball. Since the learned constraint is always added to handcrafted constraints that may already bound the action set, the argument as written does not convincingly justify the quadratic constraint choice.
  3. [Sec. 4, Table 2] The reward scales differ across scenarios (max roughly 1, 1, 1.5, and unknown in Sensor Coverage), and some baselines are not applicable to all scenarios (RVO is N/A in Connectivity and Sensor Coverage). It would be clearer to report normalized rewards per scenario or to provide a detailed per-scenario statistical comparison, rather than an unweighted average of relative improvements.
  4. [Sec. 4.1 and Fig. 3c] The collision-penalty comparison during training shows ReCoDe at -0.0001 and Online CBF at -0.06, both near zero. The text says ReCoDe 'preserves safety' but the metric is a reward penalty, not a formal safety certificate; consider reporting raw collision rates or safety violations as well.
  5. [Sec. 4.2 (Ablation) ] The ablation in Fig. 3f uses 4 random seeds; this is small, and the number of seeds should be stated in the main text or figure caption rather than only in the Appendix.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: the empirical comparison and the constructive tracking result are self-contained, and the theoretical caveat in Proposition 3.2 is an acknowledged assumption rather than a circular reduction.

full rationale

ReCoDe's central empirical claim — an 18% average reward improvement and 5% sample complexity relative to baselines — is an experimental result in which the policy is trained and evaluated on the same task reward; this is standard machine-learning practice, not a circular prediction. Proposition 3.1 is constructive rather than circular: it shows that by choosing a(t)=u*(t) and b(t)=epsilon, the QCQP can force the solver to track any strictly feasible safe trajectory, establishing expressiveness of the constraint parameterization without fitting or renaming any input. Proposition 3.2 is a conditional theorem whose hypothesis explicitly states that a weighted combination c1 Ql_i - c2 J_i approximates Q*, and whose conclusion follows from the inequalities in Appendix E; while this assumption is strong and unverified, the paper openly states that c1 and c2 are unknown and that the proposition 'does not directly imply a strategy,' so it is an acknowledged theoretical limitation rather than a self-definitional reduction. Self-citations such as Online-CBF [5] and the RoboMaster/BenchMARL/VMAS infrastructure [50,51,52] are used as baselines or implementation references, not as load-bearing justification for the paper's central claims. The evaluation protocol using the best six consecutive training steps is statistically questionable, but it is a correctness concern, not circularity. Overall, no step in the derivation chain reduces by construction to its own inputs.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the feasibility and convexity of the augmented QP, and on the ad hoc mixture assumption in Proposition 3.2. No new physical entities are introduced. The slack penalty and handcrafted controller constants are manually chosen and unreported, which limits reproducibility.

free parameters (2)
  • Slack penalty lambda_0 for learned constraints = not reported (set to 'sufficiently large' / heavily penalized)
    Used in Eq. (1) and required to be large for Propositions 3.1 and 3.2; the specific value is not given, so the exact behavior of the solver is not reproducible.
  • Handcrafted controller constants (e.g., CBF gain k, safety distance d_min, velocity limit M) = not reported
    These define the baseline QP in Appendix F and are manually calibrated (Appendix G); they affect baseline strength but the paper does not provide their values.
assumptions (3)
  • domain assumption The optimization problem (1) is a convex QCQP with strictly convex objective, non-empty convex feasible set for every observation and parameter vector, and a unique, continuously differentiable solution mapping.
    Stated in Section 2; needed for the solver to be tractable and for Propositions 3.1 and 3.2 to hold. If infeasible or nonconvex, the safety and tracking guarantees break.
  • domain assumption The desired trajectory in Proposition 3.1 is strictly feasible under the handcrafted constraints without slack (there exists eta > 0 such that B_eta(u*(t)) is inside the feasible set at each t).
    Used in the proof of Prop 3.1 to ensure that a ball around u*(t) remains feasible, so that the learned constraint can force tracking.
  • ad hoc to paper For Proposition 3.2, there exist positive constants c1, c2 such that c1 Q^l_i - c2 J_i approximates Q* within epsilon on the ball B_r(a(o)), with the critic gradient bounded by delta1 and a direction along which -J_i has directional derivative at least delta2 > delta1.
    This is a constructed 'mixture' assumption specifically designed to yield the conclusion that mixing expert and learned knowledge helps. It is not derived from first principles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReCoDe: Reinforcement Learning-based Dynamic Constraint Design for Multi-Agent Coordination." pith.science (2026). https://pith.science/paper/HVXHLRDO

@misc{pith2026250719151,
  author       = {Pith},
  title        = {Pith review of: ReCoDe: Reinforcement Learning-based Dynamic Constraint Design for Multi-Agent Coordination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HVXHLRDO}},
  note         = {Machine review of arXiv:2507.19151}
}
read the original abstract

Constraint-based optimization is a cornerstone of robotics, enabling the design of controllers that reliably encode task and safety requirements such as collision avoidance or formation adherence. However, handcrafted constraints can fail in multi-agent settings that demand complex coordination. We introduce ReCoDe--Reinforcement-based Constraint Design--a decentralized, hybrid framework that merges the reliability of optimization-based controllers with the adaptability of multi-agent reinforcement learning. Rather than discarding expert controllers, ReCoDe improves them by learning additional, dynamic constraints that capture subtler behaviors, for example, by constraining agent movements to prevent congestion in cluttered scenarios. Through local communication, agents collectively constrain their allowed actions to coordinate more effectively under changing conditions. In this work, we focus on applications of ReCoDe to multi-agent navigation tasks requiring intricate, context-based movements and consensus, where we show that it outperforms purely handcrafted controllers, other hybrid approaches, and standard MARL baselines. We give empirical (real robot) and theoretical evidence that retaining a user-defined controller, even when it is imperfect, is more efficient than learning from scratch, especially because ReCoDe can dynamically change the degree to which it relies on this controller.

Figures

Figures reproduced from arXiv: 2507.19151 by the authors.

Figure 1
Figure 1. Left: An overview of the proposed ReCoDe method. A GNN policy A aggregates the encoded observations of neighboring agents within a visibility range RN and generates constraint parameters θ that influence the feasible set U of an optimization-based controller B . Right: Real-robot position-swap in a 90 cm￾wide, 6.4 m-long corridor. Top: baseline QP controller dead-locks. Bottom: the same controller augmented with ReC… view at source ↗
Figure 2
Figure 2. Experimental scenarios and results. Leftmost column: initial conditions; middle: a possible deadlock [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. (a) Complexity vs. Reward in Waypoint Navigation: ReCoDe consistently outperforms both pure MARL and the handcrafted controller across different agent radii, demonstrating robustness in both high- and low-task complexities. Shaded regions indicate standard deviations. (b) Sample Efficiency: In Waypoint Nav￾igation with agent radius = 0.1, ReCoDe quickly converges to near-optimal reward, whereas pure MARL re￾mains su… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual depiction of the Sensor Coverage experiment. Sensors (the Os) increase the reward by ap￾proaching their color-matched Point of Interest (the Xs) but are constrained by the need to maintain formation with other sensors. Pure MARL have a difficulty optimizing this…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 51 canonical work pages

  1. [1]

    Blumenkamp, A

    J. Blumenkamp, A. Shankar, M. Bettini, J. Bird, and A. Prorok. The Cambridge RoboMas- ter: An Agile Multi-Robot Research Platform. InInternational Symposium on Distributed Autonomous Robotic Systems (DARS), 2024

    work page 2024

  2. [2]

    Alonso-Mora, T

    J. Alonso-Mora, T. Naegeli, R. Siegwart, and P. Beardsley. Collision avoidance for aerial vehicles in multi-agent scenarios.Autonomous Robots, 39:101–121, 2015

    work page 2015

  3. [3]

    Merkt, V

    W. Merkt, V . Ivan, and S. Vijayakumar. Continuous-time collision avoidance for trajectory optimization in dynamic environments. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019

    work page 2019

  4. [4]

    Gregory.Constrained optimization in the calculus of variations and optimal control theory

    J. Gregory.Constrained optimization in the calculus of variations and optimal control theory. Chapman and Hall/CRC, 2018

    work page 2018

  5. [5]

    Z. Gao, G. Yang, and A. Prorok. Online control barrier functions for decentralized multi-agent navigation. InIEEE International Symposium on Multi-Robot and Multi-Agent Systems (MRS), 2023

    work page 2023

  6. [6]

    Busoniu, R

    L. Busoniu, R. Babuska, and B. De Schutter. A comprehensive survey of multiagent reinforce- ment learning.IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), 38(2):156–172, 2008

    work page 2008

  7. [7]

    Canese, G

    L. Canese, G. C. Cardarilli, L. Di Nunzio, R. Fazzolari, D. Giardino, M. Re, and S. Span `o. Multi-agent reinforcement learning: A review of challenges and applications.Applied Sci- ences, 11(11):4948, 2021. 9

    work page 2021

  8. [8]

    Gronauer and K

    S. Gronauer and K. Diepold. Multi-agent deep reinforcement learning: a survey.Artificial Intelligence Review, 55(2):895–943, 2022

    work page 2022

Show all 54 references
  1. [9]

    Alshiekh, R

    M. Alshiekh, R. Bloem, R. Ehlers, B. K ¨onighofer, S. Niekum, and U. Topcu. Safe reinforce- ment learning via shielding. InProceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018

    2018

  2. [10]

    A. R., E. A., Y . S., and D. S. Actor-Critic Model Predictive Control: Differentiable Optimiza- tion meets Reinforcement Learning, 2024

    2024

  3. [11]

    Van den Berg, M

    J. Van den Berg, M. Lin, and D. Manocha. Reciprocal velocity obstacles for real-time multi- agent navigation. In2008 IEEE international conference on robotics and automation, pages 1928–1935. Ieee, 2008

    1928

  4. [12]

    M. G. Earl and R. D’Andrea. Modeling and control of a multi-agent system using mixed integer linear programming. InIEEE Conference on Decision and Control (CDC), 2002

    2002

  5. [13]

    Fallahi, J

    A. Fallahi, J. M. Rosenberger, V . C. P. Chen, W. Lee, and S. Wang. Linear programming for multi-agent demand response.IEEE Access, 7:181479–181490, 2019

    2019

  6. [14]

    Nocedal and S

    J. Nocedal and S. J. Wright. Quadratic programming.Numerical Optimization, pages 448–492, 2006

    2006

  7. [15]

    Nguyen and K

    Q. Nguyen and K. Sreenath. Exponential control barrier functions for enforcing high relative- degree safety-critical constraints. InIEEE American Control Conference (ACC), 2016

    2016

  8. [16]

    N. Q. H. Tran, I. Prodan, and L. Lef `evre. Nonlinear optimization for multi-agent motion planning in a multi-obstacle environment. InIEEE International Conference on System Theory, Control and Computing (ICSTCC), 2017

    2017

  9. [17]

    T. Chu, J. Wang, L. Codec `a, and Z. Li. Multi-agent deep reinforcement learning for large- scale traffic signal control.IEEE Transactions on Intelligent Transportation Systems, 21(3): 1086–1095, 2019

    2019

  10. [18]

    Xue and W

    Y . Xue and W. Chen. Multi-agent deep reinforcement learning for UA Vs navigation in un- known complex environment.IEEE Transactions on Intelligent Vehicles, 2023

    2023

  11. [19]

    Ning and L

    Z. Ning and L. Xie. A survey on multi-agent reinforcement learning and its application.Jour- nal of Automation and Intelligence, 2024

    2024

  12. [20]

    C. Amato. An introduction to centralized training for decentralized execution in cooperative multi-agent reinforcement learning.arXiv preprint arXiv:2409.03052, 2024

    2024 arXiv

  13. [21]

    A. V . Fiacco. Introduction to sensitivity and stability analysis in non linear programming. 1983

    1983

  14. [22]

    M. S. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret. Applications of second-order cone programming.Linear algebra and its applications, 284(1-3):193–228, 1998

    1998

  15. [23]

    Nayak, K

    S. Nayak, K. Choi, W. Ding, S. Dolan, K. Gopalakrishnan, and H. Balakrishnan. Scalable multi-agent reinforcement learning through intelligent information aggregation. InInterna- tional Conference on Machine Learning, pages 25817–25833. PMLR, 2023

    2023

  16. [24]

    Amos and J

    B. Amos and J. Z. Kolter. Optnet: Differentiable optimization as a layer in neural networks. InInternational Conference on Machine Learning (ICML), 2017

    2017

  17. [25]

    Blondel, Q

    M. Blondel, Q. Berthet, M. Cuturi, R. Frostig, S. Hoyer, F. Llinares-L ´opez, F.and Pedregosa, and J. Vert. Efficient and Modular Implicit Differentiation. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. 10

    2022

  18. [26]

    Agrawal, B

    A. Agrawal, B. Amos, S. Barratt, S. Boyd, S. Diamond, and Z. Kolter. Differentiable Convex Optimization Layers. InAdvances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  19. [27]

    S. Dong. methods for constrained optimization.Massachusetts Institute of Technology, 2006

    2006

  20. [28]

    Zhang, F

    J. Zhang, F. Lin, S. Ding, and W. Xing. Linear Programming-Based Consensus of Positive Continuous-Time Multi-Agent Systems.IEEE/CAA Journal of Automatica Sinica, 11(12): 2519–2521, 2024

    2024

  21. [29]

    Motee and A

    N. Motee and A. Jadbabaie. Distributed multi-parametric quadratic programming.IEEE Trans- actions on Automatic Control, 54(10):2279–2289, 2009

    2009

  22. [30]

    M. Endo, T. Ibuki, and M. Sampei. Collision-free formation control for quadrotor networks based on distributed quadratic programs. InIEEE American Control Conference (ACC), 2019

    2019

  23. [31]

    Romero, Y

    A. Romero, Y . Song, and D. Scaramuzza. Actor-critic model predictive control. InIEEE International Conference on Robotics and Automation (ICRA), 2024

    2024

  24. [32]

    Sun and C

    X. Sun and C. G. Cassandras. Optimal dynamic formation control of multi-agent systems in constrained environments.Automatica, 73:169–179, 2016

    2016

  25. [33]

    Chai and J

    J. Chai and J. K. Hodgins. Constraint-based motion optimization using a statistical dynamic model. InACM SIGGRAPH papers, pages 8–es. 2007

    2007

  26. [34]

    J. Lin, N. Somani, B. Hu, M. Rickert, and A. Knoll. An efficient and time-optimal trajectory generation approach for waypoints under kinematic constraints and error bounds. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018

    2018

  27. [35]

    Mu ˜noyerro, A

    A. Mu ˜noyerro, A. Hern ´andez, M. Urizar, and O. Altuzarra. A general automatic method for mechanism optimization based on kinematic constraints and analytical Jacobian matrix.Pro- ceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineer- in...

    2023

  28. [36]

    Gonz ´alez-Briones, F

    A. Gonz ´alez-Briones, F. De La Prieta, M. S. Mohamad, S. Omatu, and J. M. Corchado. Multi- agent systems applications in energy optimization problems: A state-of-the-art review.Ener- gies, 11(8):1928, 2018

    1928

  29. [37]

    Gao and A

    Z. Gao and A. Prorok. Environment optimization for multi-agent navigation. InIEEE Inter- national Conference on Robotics and Automation (ICRA), 2023

    2023

  30. [38]

    Gao and A

    Z. Gao and A. Prorok. Constrained environment optimization for prioritized multi-agent navi- gation.IEEE Open Journal of Control Systems, 2023

    2023

  31. [39]

    Kornienko, O

    S. Kornienko, O. Kornienko, and J. Priese. Application of multi-agent planning to the assign- ment problem.Computers in Industry, 54(3):273–290, 2004

    2004

  32. [40]

    Nedic, A

    A. Nedic, A. Ozdaglar, and P. A. Parrilo. Constrained consensus and optimization in multi- agent networks.IEEE Transactions on Automatic Control, 55(4):922–938, 2010

    2010

  33. [41]

    Zheng and L

    X. Zheng and L. Wang. A multi-agent optimization algorithm for resource constrained project scheduling problem.Expert Systems with Applications, 42(15-16):6039–6049, 2015

    2015

  34. [42]

    Bus ¸oniu, R

    L. Bus ¸oniu, R. Babuˇska, and B. De Schutter. Multi-agent reinforcement learning: An overview. Innovations in Multi-Agent Systems and Applications-1, pages 183–221, 2010

    2010

  35. [43]

    Z. Gao, G. Yang, and A. Prorok. Co-Optimization of Environment and Policies for Decentral- ized Multi-Agent Navigation.arXiv preprint arXiv:2403.14583, 2024

    2024 arXiv

  36. [44]

    Damadam, M

    S. Damadam, M. Zourbakhsh, R. Javidan, and A. Faroughi. An intelligent IoT based traffic light management system: deep reinforcement learning.Smart Cities, 5(4):1293–1311, 2022. 11

    2022

  37. [45]

    X. Pan, M. Liu, F. Zhong, Y . Yang, S.-C. Zhu, and Y . Wang. Mate: Benchmarking multi- agent reinforcement learning in distributed target coverage control. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  38. [46]

    Aydemir and A

    F. Aydemir and A. Cetin. Multi-agent dynamic area coverage based on reinforcement learning with connected agents.Computer Systems Science and Engineering, 45(1), 2023

    2023

  39. [47]

    C. S. De Witt, T. Gupta, D. Makoviichuk, V . Makoviychuk, P. H. S. Torr, M. Sun, and S. White- son. Is independent learning all you need in the starcraft multi-agent challenge?arXiv preprint arXiv:2011.09533, 2020

    2011 arXiv

  40. [48]

    Schulman, F

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

    2017 arXiv

  41. [49]

    Brody, U

    S. Brody, U. Alon, and E. Yahav. How attentive are graph attention networks? InInternational Conference on Learning Representations (ICLR), 2022

    2022

  42. [50]

    Bettini, R

    M. Bettini, R. Kortvelesy, J. Blumenkamp, and A. Prorok. VMAS: A Vectorized Multi- Agent Simulator for Collective Robot Learning.International Symposium on Distributed Au- tonomous Robotic Systems (DARS), 2022

    2022

  43. [51]

    Bettini, A

    M. Bettini, A. Prorok, and V . Moens. Benchmarl: Benchmarking multi-agent reinforcement learning.Journal of Machine Learning Research, 25(217):1–10, 2024

    2024

  44. [52]

    Blumenkamp, A

    J. Blumenkamp, A. Shankar, M. Bettini, J. Bird, and A. Prorok. The Cambridge RoboMaster: An Agile Multi-Robot Research Platform. InIEEE International Symposium on Distributed Robotic Systems (DARS), 2024

    2024

  45. [53]

    Shankar, S

    A. Shankar, S. Elbaum, and C. Detweiler. Freyja: A full multirotor system for agile & precise outdoor flights. InIEEE International Conference on Robotics and Automation (ICRA), 2021. 12 Appendix A Further Related Work on Constrained Optimization and Multi-Agent RL Constrained...

    2021

  46. [54]

    all or nothing

    This kind of objective is suboptimal since it pulls agents in different directions (due to having different goals), but they must stay connected, hence the entire cloud of agents can get stuck in a deadlock. However, it is the best we could find for a quadratic constrained opt...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.