Pith. sign in

REVIEW 4 major objections 5 minor 75 references

Symmetries-enhanced Multi-Agent Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Any control-affine system can be embedded in a symmetric one, so optimal multi-agent policies factor into a G-equivariant policy plus a learned symmetry-breaking map.

desk verdict Useful practical recipe for equivariant MARL with broken symmetries, but the theory and experiments are not as connected as claimed; still deserves a careful referee. read the letter →

arxiv 2501.01136 v2 pith:6YXMFRKG submitted 2025-01-02 cs.RO cs.AIcs.LGcs.MAmath.RT

Nikolaos Bousias , Stefanos Pertigkiozoglou , Kostas Daniilidis , George Pappas This is my paper · ORCID
classification cs.ROcs.AIcs.LGcs.MAmath.RT
keywords multi-agentreinforcementlearningequivariantpoliciesextrinsicsymmetriesgroupcanonicalizationgraphtransformerswarmroboticsquadrotorcontrolzero-shotscalability
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

Symmetry-based reinforcement learning usually assumes the robot dynamics actually have the symmetry the policy encodes. This paper argues that assumption is unnecessary: any control-affine dynamical system can be embedded in an associated system that is equivariant under a Lie group G by enlarging the input space to the span of pushforwards of the dynamics, and the optimal policy of the original problem is recovered by composing a G-equivariant policy with a symmetry-breaking map. The construction turns a task with a G-invariant reward into one to which equivariant policy learning applies, even when the underlying dynamics break G. To implement this, the paper introduces a Group Equivariant Graphormer that achieves equivariance by canonicalizing features rather than constraining the network architecture. In quadrotor-swarm simulations with embedded SE(3) symmetry, the resulting policies collide less and transfer to larger swarms better than non-equivariant baselines.

What carries the argument

The load-bearing objects are the push-forward group action on vector fields, $d^*\phi_g f(x,u):=d\phi_g f(\phi_{g^{-1}}(x),u)$, and the extended input space $\hat U:=\operatorname{span}\{d^*\phi_g f(x,u)\mid u\in U,\ g\in G\}$. Theorem 2 shows the associated dynamics $F(x,\hat u)=\sum_j d^*\phi_{g_j} f(x,u_j)$ are $G$-equivariant, and Proposition 2 decomposes the original optimal policy as $\pi^*=h_U\circ\hat\pi^*$ with $\hat\pi^*$ $G$-equivariant and $h_U$ a symmetry-breaking projection. To implement this without specialized layers, Lemma 2 canonicalizes equivariance: any equivariant map $f(g,x)$ equals $\psi_g[h(\phi_{g^{-1}}x)]$ for an unconstrained function $h$. The Group Equivariant Graphormer applies that canonicalization inside each transformer update as $M_v^{\mathrm{eq}}(g_v,F^{(l)})=\phi_{g_v}^{(l+1)}[M_v(\phi_{g_v^{-1}}^{(l)}[F^{(l)}])]$, turning an ordinary graph transformer into a group-equivariant policy network.

What would settle it

Measure the equivariance error of the trained equivariant head on held-out inputs: for sampled states $s$ and group elements $g$, compute $\|\hat\pi_\theta(\phi_g(s))-\psi_g(\hat\pi_\theta(s))\|$. If the error is large while the full policy still shows the reported collision and success gains, the benefit is not coming from the head realizing the claimed equivariant lift, which would undercut the practical reading of Proposition 2.

Watch

Extended reading notes

Core claim

The paper's central claim is that a lack of intrinsic symmetry in the dynamics is not a barrier to equivariant policy learning. Theorem 2 states that any control-affine system $f:X\times U\to \mathfrak{X}(X)$ can be embedded in an associated system $F$ on $\hat U\times X$ that is equivariant under any Lie group $G$ compatible with the state manifold, by defining $\hat U$ as the span of the pushforwards $d^*\phi_g f(x,u)$. Proposition 2 then shows the optimal policy of the original problem is exactly $h_U\circ\hat\pi^*$, where $\hat\pi^*$ is a $G$-equivariant optimal policy for the associated system and $h_U$ is a smooth idempotent map back to $U$. The paper therefore claims that any multi-agent RL problem whose reward is $G$-invariant can be solved by learning a $G$-equivariant policy followed by a symmetry-breaking head, and supports this with quadrotor-swarm experiments in which the embedded SE(3)-equivariant policies reduce collisions, reach 100 percent success in some scenarios, and keep high success rates when the trained policy is transferred zero-shot to swarms of 64 and 128 agents.

Load-bearing premise

Everything rests on the learned equivariant half of the policy finding the correct symmetric solution even though it is trained only through the full policy and never directly on the extended symmetric problem; the paper explicitly says this convergence is not guaranteed.

Editorial extensions

If this is right

  • Any control-affine multi-agent RL problem with a $G$-invariant reward can be solved by a $G$-equivariant policy composed with a learned symmetry-breaking head, so equivariant learning no longer requires the dynamics themselves to admit $G$.
  • The canonicalization lemma lets existing non-equivariant architectures be made equivariant by transforming their inputs and outputs, avoiding the optimization difficulties of specialized equivariant layers.
  • On the paper's eight-quadrotor benchmarks, the SE(3)-enhanced policies reduce per-scenario collision counts to 0-0.55 while baselines such as K-Attention and rMAPPO rise to 1-3 in the same scenarios.
  • Zero-shot scalability: policies trained on eight quadrotors keep roughly 97.5 percent success with 64 drones and 84.6 percent with 128 drones, while the baselines fall below 65 percent at 128.
  • The choice of symmetry group becomes the main remaining design decision, since any group compatible with the state manifold and with the reward can be embedded, discrete or continuous.

Reading between the lines

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

  • Beyond the reported quadrotor experiments, the embedding theorem is stated for any control-affine system, so the same two-part policy should transfer to other homogeneous robot teams whose task reward can be written in a frame-invariant way.
  • Because the equivariance is achieved by canonicalizing inputs and outputs rather than by new layer types, a testable prediction is that wrapping any strong permutation-equivariant MARL backbone with the same SE(3) canonicalization would reproduce most of the reported gains.
  • The paper leaves symmetry-group selection to future work; in practice the choice of G becomes a hyperparameter that trades how much structure is injected against the effective dimension of the lifted action space and how well the learned head can break symmetry.
  • Since the paper concedes there is no guarantee the learned equivariant head reaches the lifted optimum, part of the measured benefit may come from the equivariant structure acting as regularization or implicit data augmentation; measuring the equivariance error of the learned head directly would separate the two mechanisms.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper addresses MARL for systems whose dynamics do not exhibit the symmetries of the task. The authors propose to embed a control-affine dynamical system into an associated equivariant system on an extended input space Û, prove that the original optimal policy can be recovered as h_U ∘ π̂* with π̂* G-equivariant when the reward is G-invariant, and introduce a Group Equivariant Graphormer that achieves equivariance through canonicalization of per-node frames. The method is evaluated on simulated Crazyflie swarms, comparing SE(3)-enhanced policies against K-Attention, rMAPPO, InfoMARL, and K-DeepSets baselines in tasks such as formation flight and evasion-pursuit, with additional zero-shot scalability tests from 16 to 128 agents. The paper claims improved collision rates, success rates, and generalization from embedding extrinsic SE(3) symmetry.

Significance. If fully substantiated, the paper would make a useful contribution to equivariant MARL by relaxing the usual requirement that the dynamics themselves be symmetric. The modular canonicalization lemma (Lemma 2) and the Group Equivariant Graphormer are sensible design elements, and the experimental setup covers a realistic multi-quadrotor domain with several scenarios and swarm sizes. The paper does not release code or seeds and its proofs are not machine-checked; its strongest assets are the clear formulation of the symmetry-breaking decomposition and the breadth of the simulation study. However, the proof of Proposition 2 is incomplete, the implemented policy does not use the extended input space of the theory, and the experimental tables lack variance measures, so the central claim is currently under-supported.

major comments (4)
  1. [Appendix I, second 'Proof [Theorem 2]' (proof of Proposition 2)] The displayed derivation ends with h_U(π̂*(φ_g(x))) = h_U(ψ_g(π̂*(x))), but Proposition 2 requires the unprojected equality π̂*(φ_g(x)) = ψ_g(π̂*(x)). Because h_U is a non-injective surjection from Û to U, equality after applying h_U does not imply equality in Û. The chain also uses h_U(h_U(û)) = h_U(û) and ψ_g^{-1}(h_U(û)) at points where the arguments are not shown to lie in the domain of h_U; these steps are not justified. A repaired argument would need to construct π̂* explicitly on group orbits and define h_U consistently on the lifted space; as written, the existence of a G-equivariant optimal lifted policy is not proven.
  2. [Section 4 and Section 5.2] The theoretical result guarantees a decomposition π* = h_U ∘ π̂* on an extended input space Û, but the implemented policy is π_θ = h_θp ∘ π̂_θe, where π̂_θe maps observations to learned feature vectors and h_θp is an unconstrained MLP; the extended input space Û is never constructed. Section 4 explicitly concedes that 'there is no guarantee that π̂θ → π̂*'. Consequently the reported experimental gains in Section 6 cannot be attributed to the extrinsic-symmetry embedding of Theorem 2; they may be due to the canonicalization mechanism or to the Graphormer architecture itself. This is a load-bearing gap between the theoretical and empirical parts of the paper.
  3. [Section 4 (finite-dimensionality assumption)] The universal-approximation step in Section 4 relies on the assumption that the extended input space Û is finite-dimensional, but the paper does not prove this for any example, including the SE(3) quadrotor swarm. The text only says that the dimension 'in practice ... depends on the complexity of the dynamical system and the selected symmetry group.' Since the reward is SE(3)-invariant while the quadrotor dynamics break SE(3), it is unclear whether the associated equivariant system F has a finite-dimensional input space; without a concrete construction this part of the theory is an assumption rather than a theorem.
  4. [Section 6, Tables 1 and 2] The empirical evaluation reports point estimates averaged over 50 episodes with no error bars, no number of seeds, and no statistical significance tests, and the code is not released. Moreover, the comparison does not isolate the claimed mechanism: SE(3)-Graphormer differs from K-Attention in architecture, canonicalization, and learned head h_θp, so the gains cannot be unambiguously assigned to extrinsic symmetry embedding. An ablation using the same graph transformer with and without the canonicalized frame, or with a fixed identity frame, would be needed to support the attribution in the current text.
minor comments (5)
  1. [Section 2.1, Definition 2] In Definition 2 the codomain action is written as ψ_g : X → X, but for a function f : X → Y the action on the codomain should be ψ_g : Y → Y; this is likely a typographical error and should be corrected.
  2. [Sections 2.2 and 3, notation for group actions] The order of arguments in the group action is inconsistent: Section 2.2 writes ϕ : G × X → X and Definition 1 writes ϕ_g : X → X, while Definition 3 and the problem statement use ϕ : X × G → X; these should be unified.
  3. [Throughout] The text contains unresolved placeholders such as 'supplemental material??' in Sections 3, 4, and 6, and typographical errors such as 'arcitectures', 'Graphomer', and 'Shur' for Schur; these should be fixed before publication.
  4. [Appendix I, proof labels] The proof of Proposition 2 is labeled 'Proof [Theorem 2]', which is the same label as the preceding proof; the duplicate labels make it difficult to identify which statement is being proved.
  5. [Figure 2] The architecture schematic in Figure 2 contains garbled characters and unexplained symbols in the caption, making the diagram difficult to interpret; a cleaner figure with a proper caption is needed.

Circularity Check

1 steps flagged · score 3.0 of 10

The extrinsic-symmetry embedding theorem is definitional (F is equivariant because Û is defined as its group orbit), but the empirical gains are independently measured and not forced by that theorem.

  1. self definitional [Section 4, Theorem 2 and the construction following it]
    "For the extended input vector space ˆU := span{d∗ϕgf (x, u) | u ∈ U, g∈ G}, the associated system dynamics F : ˆU × X →X(X ) are equivariant with respect to actions induced by elements of the Lie group G."

    The paper defines F(x, ˆu) := Σ_j α_j d∗ϕ_{g_j} f(x, u_j) on the span of the pushforwards. Equivariance of F is then exactly the statement that this span is closed under the pushforward action, which is how ˆU was defined in the first place. Theorem 2 therefore restates the definition of ˆU rather than deriving an independent property of control-affine systems. This is a definitional construction, not an empirical prediction. The later practical claims do not reduce to this theorem because the trained h_θ ∘ π̂_θe never constructs ˆU, and Tables 1 and 2 are separate measurements.

full rationale

Most of the paper is a standard application of equivariant control and RL: Theorem 1 derives policy equivariance from equivariant dynamics and rewards, and Lemma 2 is a proved canonicalization identity. The one definitional step is Theorem 2: the 'associated equivariant system' is equivariant by construction because its input space is defined as the closure of the original dynamics under the group action. That step is tautological but is not what produces the empirical numbers; the trained policy h_θ ∘ π̂_θe never uses the extended space Û, and Section 4 explicitly admits 'there is no guarantee that π̂θ → π̂*'. Tables 1 and 2 are independent measurements of the actual architecture. A genuine proof gap accompanies Proposition 2: its appendix proof only establishes h_U(π̂*(φ_g(x))) = h_U(ψ_g(π̂*(x))) and then asserts the equivariance of π̂* itself; this is an omitted proof, not a circularity. Self-citations in the paper are incidental and not load-bearing. The score 3 reflects one definitional theoretical result plus independent empirical content that does not reduce to that result.

Assumptions & free parameters 2 free parameters · 6 assumptions · 1 invented entities

The central construction introduces an extended input space and an associated equivariant system F that are not used in the implementation. The practical results depend on hand-chosen reward coefficients and learned network weights, plus several structural assumptions about the reward and the dynamics.

free parameters (2)
  • Reward coefficients c1-c5 = c1=0.5*dt, c2=5, c3=5*dt, c4=0.01*dt, c5=0.05*dt
    Hand-chosen constants in the reward function (Appendix III); the learned policies depend on these values, and the reported collision and success metrics are defined relative to them.
  • Learned symmetry-breaking projection h_theta = MLP weights
    The framework requires learning h_U, the idempotent surjective projection from extended to original inputs; in experiments this is a trained neural network (the 'projection MLP'), so the central decomposition pi* = h_U ∘ pi_hat* is not parameter-free.
assumptions (6)
  • domain assumption The reward function L is invariant under the chosen group G (Definition 3, condition 1).
    The entire framework requires condition 1 to hold. For the quadrotor experiments, SE(3)-invariance of the reward is asserted, but no obstacle term appears in the reward, so the assumption is untested if obstacles are present.
  • domain assumption The neighborhood structure is G-invariant (Assumption 1).
    Required for Theorem 1 and for the local graph features to transform equivariantly; holds only if neighborhoods are defined by Euclidean distance.
  • standard math The value function is differentiable and the HJB equation applies (Appendix I proof of Theorem 1).
    The proof of optimal policy equivariance uses the HJB equation, which requires smoothness of V* and unconstrained control; this is a regularity assumption.
  • domain assumption Universal approximation holds for the equivariant network class (Section 4).
    The paper invokes universal approximation to claim pi_theta can approximate pi*, which is plausible for MLPs but not proven for the specific canonicalized architecture.
  • domain assumption The dynamics are control-affine (Theorem 2).
    The pushforward construction uses control-affine structure; the quadrotor model used in experiments is not shown to be control-affine in the stated form.
  • ad hoc to paper The extended input space is finite-dimensional in practice (Section 4).
    The framework acknowledges U_hat may be infinite-dimensional; the experiments avoid the issue by not using U_hat at all, so this axiom is silently discharged.
invented entities (1)
  • Associated equivariant dynamical system F (extended input space U_hat)
    purpose: To embed non-equivariant dynamics into a G-equivariant system so that Theorem 1 can be applied.
    F is defined by construction as the span of pushforwards of the original dynamics; it is a mathematical device with no falsifiable handle outside the paper, and the experiments never instantiate it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Symmetries-enhanced Multi-Agent Reinforcement Learning." pith.science (2026). https://pith.science/paper/6YXMFRKG

@misc{pith2026250101136,
  author       = {Pith},
  title        = {Pith review of: Symmetries-enhanced Multi-Agent Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6YXMFRKG}},
  note         = {Machine review of arXiv:2501.01136}
}
read the original abstract

Multi-agent reinforcement learning has emerged as a powerful framework for enabling agents to learn complex, coordinated behaviors but faces persistent challenges regarding its generalization, scalability and sample efficiency. Recent advancements have sought to alleviate those issues by embedding intrinsic symmetries of the systems in the policy. Yet, most dynamical systems exhibit little to no symmetries to exploit. This paper presents a novel framework for embedding extrinsic symmetries in multi-agent system dynamics that enables the use of symmetry-enhanced methods to address systems with insufficient intrinsic symmetries, expanding the scope of equivariant learning to a wide variety of MARL problems. Central to our framework is the Group Equivariant Graphormer, a group-modular architecture specifically designed for distributed swarming tasks. Extensive experiments on a swarm of symmetry-breaking quadrotors validate the effectiveness of our approach, showcasing its potential for improved generalization and zero-shot scalability. Our method achieves significant reductions in collision rates and enhances task success rates across a diverse range of scenarios and varying swarm sizes.

Figures

Figures reproduced from arXiv: 2501.01136 by the authors.

Figure 1
Figure 1. Instances of the swarm in various scenarios. [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 3
Figure 3. Instances of the swarm in various scenarios [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 52 canonical work pages

  1. [1]

    11em plus .33em minus .07em @technote 4000 4000 100 4000 4000 500 `\.=1000 = #1 #1 #1 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEauthorblockAstyle \@IEEEauthordefaulttextstyle \@IEEEauthorblockconfadjspace -0.25em \@IEEEauthorblockNtopspace 0.0ex \@IEEEauthorblockAtopspace 0.0ex \@IEEEauthorblockNinterlinespace 2.6ex \@IEEEauthorblockAinte...

    arXiv 2023

  2. [2]

    On learning symmetric locomotion

    Farzad Abdolhosseini, Hung Yu Ling, Zhaoming Xie, Xue Bin Peng, and Michiel Van de Panne. On learning symmetric locomotion. In Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games, pages 1--10, 2019

    work page 2019

  3. [3]

    Deep reinforcement learning meets graph neural networks: Exploring a routing optimization use case

    Paul Almasan, José Suárez-Varela, Krzysztof Rusek, Pere Barlet-Ros, and Albert Cabellos-Aparicio. Deep reinforcement learning meets graph neural networks: Exploring a routing optimization use case. Computer Communications, 196: 0 184–194, December 2022. ISSN 0140-3664. doi:10.1016/j.comcom.2022.09.029. URL http://dx.doi.org/10.1016/j.comcom.2022.09.029

    work page doi:10.1016/j.comcom.2022.09.029 2022

  4. [4]

    Decentralized Control of Quadrotor Swarms with End-to-end Deep Reinforcement Learning

    Sumeet Batra, Zhehui Huang, Aleksei Petrenko, Tushar Kumar, Artem Molchanov, and Gaurav S. Sukhatme. Decentralized control of quadrotor swarms with end-to-end deep reinforcement learning. In 5th Conference on Robot Learning, CoRL 2021, 8-11 November 2021, London, England, UK , Proceedings of Machine Learning Research. PMLR , 2021. URL https://arxiv.org/ab...

    work page Pith review arXiv 2021

  5. [5]

    Erik J Bekkers, Sharvaree Vadgama, Rob Hesselink, Putri A Van der Linden, and David W. Romero. Fast, expressive \ mathrm\ SE \ (n)\ equivariant networks through weight-sharing in position-orientation space. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=dPHLbUqGbr

    2024

  6. [6]

    Deep coordination graphs

    Wendelin Boehmer, Vitaly Kurin, and Shimon Whiteson. Deep coordination graphs. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 980--991. PMLR, 13--18 Jul 2020. URL https://proceedings.mlr.press/v119/boehmer20a.html

    work page 2020

  7. [7]

    E(3)-equivariant actor-critic methods for cooperative multi-agent reinforcement learning

    Dingyang Chen and Qi Zhang. E(3)-equivariant actor-critic methods for cooperative multi-agent reinforcement learning. arXiv preprint arXiv:2308.11842, 2023

    arXiv 2023

  8. [8]

    Communication-efficient actor-critic methods for homogeneous markov games

    Dingyang Chen, Yile Li, and Qi Zhang. Communication-efficient actor-critic methods for homogeneous markov games. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=xy_2w3J3kH

    work page 2022

Show all 75 references
  1. [9]

    Subequivariant graph reinforcement learning in 3 D environments

    Runfa Chen, Jiaqi Han, Fuchun Sun, and Wenbing Huang. Subequivariant graph reinforcement learning in 3 D environments. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conf...

    2023

  2. [10]

    Training deep convolutional neural networks to play go

    Christopher Clark and Amos Storkey. Training deep convolutional neural networks to play go. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 1766--1774, L...

    2015

  3. [11]

    Gauge equivariant mesh cnns: Anisotropic convolutions on geometric graphs, 2021

    Pim de Haan, Maurice Weiler, Taco Cohen, and Max Welling. Gauge equivariant mesh cnns: Anisotropic convolutions on geometric graphs, 2021. URL https://arxiv.org/abs/2003.05425

    2021 arXiv

  4. [12]

    Automatic symmetry discovery with lie algebra convolutional network

    Nima Dehmamy, Robin Walters, Yanchen Liu, Dashun Wang, and Rose Yu. Automatic symmetry discovery with lie algebra convolutional network. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, v...

    2021

  5. [13]

    Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J. Guibas. Vector neurons: A general framework for so(3)-equivariant networks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12180--12189, 2021. URL https://api.s...

    2021

  6. [14]

    Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data

    Marc Finzi, Samuel Stanton, Pavel Izmailov, and Andrew Gordon Wilson. Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In International Conference on Machine Learning, pages 3165--3176. PMLR, 2020

    2020

  7. [15]

    System identification of the crazyflie 2.0 nano quadrocopter

    Julian F \"o rster. System identification of the crazyflie 2.0 nano quadrocopter. 2015. URL https://api.semanticscholar.org/CorpusID:139392646

    2015

  8. [16]

    Fuchs, Daniel E

    Fabian B. Fuchs, Daniel E. Worrall, Volker Fischer, and Max Welling. Se(3)-transformers: 3d roto-translation equivariant attention networks. In Advances in Neural Information Processing Systems 34 (NeurIPS), 2020

    2020

  9. [17]

    Notes on group actions manifolds, lie groups and lie algebras

    Jean Gallier. Notes on group actions manifolds, lie groups and lie algebras. Technical report, University of Pennsylvania, 2005

    2005

  10. [18]

    Grizzle and S

    J. Grizzle and S. Marcus. Optimal control of systems possessing symmetries. IEEE Transactions on Automatic Control, 29 0 (11): 0 1037--1040, 1984. doi:10.1109/TAC.1984.1103421

    1984

  11. [19]

    Exploiting different symmetries for trajectory tracking control with application to quadrotors*

    Matthew Hampsey, Pieter van Goor , Tarek Hamel, and Robert Mahony. Exploiting different symmetries for trajectory tracking control with application to quadrotors*. IFAC-PapersOnLine, 56 0 (1): 0 132--137, 2023. ISSN 2405-8963. doi:https://doi.org/10.1016/j.ifacol.2023.02.023. ...

    2023 doi

  12. [20]

    Boosting multiagent reinforcement learning via permutation invariant and permutation equivariant networks

    Jianye HAO, Xiaotian Hao, Hangyu Mao, Weixun Wang, Yaodong Yang, Dong Li, YAN ZHENG, and Zhen Wang. Boosting multiagent reinforcement learning via permutation invariant and permutation equivariant networks. In The Eleventh International Conference on Learning Representations, ...

    2023

  13. [21]

    Gauge equivariant transformer

    Lingshen He, Yiming Dong, Yisen Wang, Dacheng Tao, and Zhouchen Lin. Gauge equivariant transformer. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 27331--27343. Curran ...

    2021

  14. [22]

    O ther-play

    Hengyuan Hu, Adam Lerer, Alex Peysakhovich, and Jakob Foerster. “ O ther-play” for zero-shot coordination. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, p...

  15. [23]

    Edge grasp network: A graph-based se(3)-invariant approach to grasp detection

    Haojie Huang, Dian Wang, Xupeng Zhu, Robin Walters, and Robert Platt. Edge grasp network: A graph-based se(3)-invariant approach to grasp detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3882--3888, 2023 a . doi:10.1109/ICRA48891.2023.10160728

    2023

  16. [24]

    Noack, Gang Hu, Shugao Liu, Yuchen Xu, and Huanhui Cao

    Junchang Huang, Weifeng Zeng, Hao Xiong, Bernd R. Noack, Gang Hu, Shugao Liu, Yuchen Xu, and Huanhui Cao. Symmetry-informed reinforcement learning and its application to low-level attitude control of quadrotors. IEEE Transactions on Artificial Intelligence, 5 0 (3): 0 1147--11...

    2024

  17. [25]

    Quadswarm: A modular multi-quadrotor simulator for deep reinforcement learning with direct thrust control

    Zhehui Huang, Sumeet Batra, Tao Chen, Rahul Krupani, Tushar Kumar, Artem Molchanov, Aleksei Petrenko, James A Preiss, Zhaojing Yang, and Gaurav S Sukhatme. Quadswarm: A modular multi-quadrotor simulator for deep reinforcement learning with direct thrust control. arXiv preprint...

    2023 arXiv

  18. [26]

    Graph convolutional reinforcement learning, 2020

    Jiechuan Jiang, Chen Dun, Tiejun Huang, and Zongqing Lu. Graph convolutional reinforcement learning, 2020. URL https://arxiv.org/abs/1810.09202

    2020 arXiv

  19. [27]

    Image augmentation is all you need: Regularizing deep reinforcement learning from pixels, 2021

    Ilya Kostrikov, Denis Yarats, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels, 2021. URL https://arxiv.org/abs/2004.13649

    2021 arXiv

  20. [28]

    Reinforcement learning with augmented data

    Misha Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages ...

    2020

  21. [29]

    HGAP : Boosting permutation invariant and permutation equivariant in multi-agent reinforcement learning via graph attention network

    Bor-Jiun Lin and Chun-Yi Lee. HGAP : Boosting permutation invariant and permutation equivariant in multi-agent reinforcement learning via graph attention network. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix ...

  22. [30]

    Invariant transform experience replay: Data augmentation for deep reinforcement learning

    Yijiong Lin, Jiancong Huang, Matthieu Zimmer, Yisheng Guan, Juan Rojas, and Paul Weng. Invariant transform experience replay: Data augmentation for deep reinforcement learning. IEEE Robotics and Automation Letters, 5 0 (4): 0 6615--6622, 2020. doi:10.1109/LRA.2020.3013937

    2020

  23. [31]

    Yeh, and Alexander G

    Iou-Jen Liu, Raymond A. Yeh, and Alexander G. Schwing. Pic: Permutation invariant critic for multi-agent deep reinforcement learning, 2019. URL https://arxiv.org/abs/1911.00025

    2019 arXiv

  24. [32]

    Enabling equivariance for arbitrary lie groups, 2022

    Lachlan Ewen MacDonald, Sameera Ramasinghe, and Simon Lucey. Enabling equivariance for arbitrary lie groups, 2022. URL https://arxiv.org/abs/2111.08251

    2022 arXiv

  25. [33]

    Equivariant filter design for kinematic systems on lie groups, 2020

    Robert Mahony and Jochen Trumpf. Equivariant filter design for kinematic systems on lie groups, 2020. URL https://arxiv.org/abs/2004.00828

    2020 arXiv

  26. [34]

    Observers for kinematic systems with symmetry

    Robert Mahony, Jochen Trumpf, and Tarek Hamel. Observers for kinematic systems with symmetry. IFAC Proceedings Volumes, 46 0 (23): 0 617--633, 2013. ISSN 1474-6670. doi:https://doi.org/10.3182/20130904-3-FR-2041.00212. URL https://www.sciencedirect.com/science/article/pii/S147...

    2013

  27. [36]

    Mahony, T

    Robert E. Mahony, T. Hamel, and Jochen Trumpf. Equivariant systems theory and observer design. ArXiv, abs/2006.08276, 2020 b . URL https://api.semanticscholar.org/CorpusID:219687199

    2006 arXiv

  28. [37]

    Exploiting symmetry for discrete-time reachability computations

    John Maidens and Murat Arcak. Exploiting symmetry for discrete-time reachability computations. IEEE Control Systems Letters, 2 0 (2): 0 213--217, 2018. doi:10.1109/LCSYS.2018.2800125

    2018

  29. [38]

    Symmetry reduction for dynamic programming, 2018

    John Maidens, Axel Barrau, Silvere Bonnabel, and Murat Arcak. Symmetry reduction for dynamic programming, 2018. URL https://arxiv.org/abs/1801.03237

    2018 arXiv

  30. [39]

    Goal-conditioned batch reinforcement learning for rotation invariant locomotion

    Aditi Mavalankar. Goal-conditioned batch reinforcement learning for rotation invariant locomotion. arXiv preprint arXiv:2004.08356, 2020

    2004 arXiv

  31. [40]

    Boosting sample efficiency and generalization in multi-agent reinforcement learning via equivariance, 2024

    Joshua McClellan, Naveed Haghani, John Winder, Furong Huang, and Pratap Tokekar. Boosting sample efficiency and generalization in multi-agent reinforcement learning via equivariance, 2024. URL https://arxiv.org/abs/2410.02581

    2024 arXiv

  32. [41]

    Lie group decompositions for equivariant neural networks

    Mircea Mironenco and Patrick Forr \'e . Lie group decompositions for equivariant neural networks. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=p34fRKp8qA

    2024

  33. [42]

    Group equivariant deep reinforcement learning, 2020

    Arnab Kumar Mondal, Pratheeksha Nair, and Kaleem Siddiqi. Group equivariant deep reinforcement learning, 2020. URL https://arxiv.org/abs/2007.03437

    2020 arXiv

  34. [43]

    Attending to graph transformers, 2023

    Luis Müller, Mikhail Galkin, Christopher Morris, and Ladislav Rampášek. Attending to graph transformers, 2023

    2023

  35. [44]

    Scalable multi-agent reinforcement learning through intelligent information aggregation

    Siddharth Nayak, Kenneth Choi, Wenqi Ding, Sydney Dolan, Karthik Gopalakrishnan, and Hamsa Balakrishnan. Scalable multi-agent reinforcement learning through intelligent information aggregation. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. J...

    2023

  36. [45]

    Policy gradient with value function approximation for collective multiagent planning

    Duc Thien Nguyen, Akshat Kumar, and Hoong Chuin Lau. Policy gradient with value function approximation for collective multiagent planning. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Pr...

    2017

  37. [46]

    Equivariant reinforcement learning under partial observability

    Hai Huu Nguyen, Andrea Baisero, David Klee, Dian Wang, Robert Platt, and Christopher Amato. Equivariant reinforcement learning under partial observability. In 7th Annual Conference on Robot Learning, 2023. URL https://openreview.net/forum?id=AnDDMQgM7-

    2023

  38. [47]

    On discrete symmetries of robotics systems: A group-theoretic and data-driven analysis

    Daniel Ordonez-Apraez, Mario Martin, Antonio Agudo, and Francesc Moreno-Noguer. On discrete symmetries of robotics systems: A group-theoretic and data-driven analysis. arXiv preprint arXiv:2302.10433, 2023

    2023 arXiv

  39. [48]

    Schoellig

    Jacopo Panerati, Hehui Zheng, SiQi Zhou, James Xu, Amanda Prorok, and Angela P. Schoellig. Learning to fly---a gym environment with pybullet physics for reinforcement learning of multi-agent quadcopter control. In 2021 IEEE/RSJ International Conference on Intelligent Robots an...

    2021

  40. [49]

    Improving equivariant model training via constraint relaxation

    Stefanos Pertigkiozoglou, Evangelos Chatzipantazis, Shubhendu Trivedi, and Kostas Daniilidis. Improving equivariant model training via constraint relaxation. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Infor...

    2024

  41. [50]

    Sukhatme, and Vladlen Koltun

    Aleksei Petrenko, Zhehui Huang, Tushar Kumar, Gaurav S. Sukhatme, and Vladlen Koltun. Sample factory: Egocentric 3d control from pixels at 100000 FPS with asynchronous reinforcement learning. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 1...

    2020

  42. [51]

    Monotonic value function factorisation for deep multi-agent reinforcement learning, 2020

    Tabish Rashid, Mikayel Samvelyan, Christian Schroeder de Witt, Gregory Farquhar, Jakob Foerster, and Shimon Whiteson. Monotonic value function factorisation for deep multi-agent reinforcement learning, 2020. URL https://arxiv.org/abs/2003.08839

    2020 arXiv

  43. [52]

    Balaraman Ravindran and Andrew G. Barto. Symmetries and model minimization in markov decision processes. 2001. URL https://api.semanticscholar.org/CorpusID:59092891

    2001

  44. [53]

    E(n) equivariant graph neural networks

    V\' ctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E(n) equivariant graph neural networks. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 9323...

    2021

  45. [54]

    Equivariant message passing for the prediction of tensorial properties and molecular spectra

    Kristof Sch \"u tt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Pr...

    2021

  46. [55]

    Gregor N. C. Simm, Robert Pinsler, Gábor Csányi, and José Miguel Hernández-Lobato. Symmetry-aware actor-critic for 3d molecular design, 2020. URL https://arxiv.org/abs/2011.12747

    2020 arXiv

  47. [56]

    So(2)-equivariant downwash models for close proximity flight

    Henry Smith, Ajay Shankar, Jennifer Gielis, Jan Blumenkamp, and Amanda Prorok. So(2)-equivariant downwash models for close proximity flight. IEEE Robotics Autom. Lett., 9 0 (2): 0 1174--1181, February 2024. URL https://doi.org/10.1109/LRA.2023.3337701

    2024

  48. [57]

    Exploiting symmetry in dynamics for model-based reinforcement learning with asymmetric rewards, 2024

    Yasin Sonmez, Neelay Junnarkar, and Murat Arcak. Exploiting symmetry in dynamics for model-based reinforcement learning with asymmetric rewards, 2024. URL https://arxiv.org/abs/2403.19024

    2024 arXiv

  49. [58]

    Leveraging symmetry in rl-based legged locomotion control

    Zhi Su, Xiaoyu Huang, Daniel Ordo \ n ez-Apraez, Yunfei Li, Zhongyu Li, Qiayuan Liao, Giulio Turrisi, Massimiliano Pontil, Claudio Semini, Yi Wu, et al. Leveraging symmetry in rl-based legged locomotion control. arXiv preprint arXiv:2403.17320, 2024

    2024 arXiv

  50. [59]

    Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018

    2018 arXiv

  51. [60]

    Graph neural networks for multi-robot active information acquisition

    Mariliza Tzes, Nikolaos Bousias, Evangelos Chatzipantazis, and George J Pappas. Graph neural networks for multi-robot active information acquisition. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3497--3503. IEEE, 2023

    2023

  52. [61]

    Coordinated deep reinforcement learners for traffic light control

    Elise Van der Pol and Frans A Oliehoek. Coordinated deep reinforcement learners for traffic light control. Proceedings of learning, inference and control of multi-agent systems (at NIPS 2016), 8: 0 21--38, 2016

    2016

  53. [62]

    Oliehoek, and Max Welling

    Elise van der Pol, Thomas Kipf, Frans A. Oliehoek, and Max Welling. Plannable approximations to mdp homomorphisms: Equivariance under actions, 2020 a . URL https://arxiv.org/abs/2002.11963

    2020 arXiv

  54. [63]

    Mdp homomorphic networks: Group symmetries in reinforcement learning

    Elise van der Pol, Daniel Worrall, Herke van Hoof, Frans Oliehoek, and Max Welling. Mdp homomorphic networks: Group symmetries in reinforcement learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Syste...

  55. [64]

    Oliehoek, and Max Welling

    Elise van der Pol, Herke van Hoof, Frans A. Oliehoek, and Max Welling. Multi-agent mdp homomorphic networks, 2022. URL https://arxiv.org/abs/2110.04495

    2022 arXiv

  56. [65]

    Policy learning in se(3) action spaces, 2020

    Dian Wang, Colin Kohler, and Robert Platt. Policy learning in se(3) action spaces, 2020. URL https://arxiv.org/abs/2010.02798

    2020 arXiv

  57. [66]

    Equivariant q learning in spatial action spaces

    Dian Wang, Robin Walters, Xupeng Zhu, and Robert Platt. Equivariant q learning in spatial action spaces. In Aleksandra Faust, David Hsu, and Gerhard Neumann, editors, Proceedings of the 5th Conference on Robot Learning, volume 164 of Proceedings of Machine Learning Research, p...

    2022

  58. [67]

    Wong, Robin Walters, and Robert Platt

    Dian Wang, Jung Yeon Park, Neel Sortur, Lawson L.S. Wong, Robin Walters, and Robert Platt. The surprising effectiveness of equivariant models in domains with latent symmetry. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net...

    2023

  59. [68]

    Koopman q-learning: Offline reinforcement learning via symmetries of dynamics

    Matthias Weissenbacher, Samarth Sinha, Animesh Garg, and Kawahara Yoshinobu. Koopman q-learning: Offline reinforcement learning via symmetries of dynamics. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Proceedings of t...

    2022

  60. [69]

    Mean field multi-agent reinforcement learning

    Yaodong Yang, Rui Luo, Minne Li, Ming Zhou, Weinan Zhang, and Jun Wang. Mean field multi-agent reinforcement learning. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning...

    2018

  61. [70]

    Equivariant reinforcement learning for quadrotor uav

    Beomyeol Yu and Taeyoung Lee. Equivariant reinforcement learning for quadrotor uav. 2023 American Control Conference (ACC), pages 2842--2847, 2022. URL https://api.semanticscholar.org/CorpusID:249375239

    2023

  62. [71]

    Equivariant reinforcement learning for quadrotor uav

    Beomyeol Yu and Taeyoung Lee. Equivariant reinforcement learning for quadrotor uav. In 2023 American Control Conference (ACC), pages 2842--2847, 2023 a . doi:10.23919/ACC55779.2023.10156379

    2023

  63. [72]

    Equivariant reinforcement learning for quadrotor uav

    Beomyeol Yu and Taeyoung Lee. Equivariant reinforcement learning for quadrotor uav. In 2023 American Control Conference (ACC), pages 2842--2847. IEEE, 2023 b

    2023

  64. [73]

    The surprising effectiveness of PPO in cooperative multi-agent games

    Chao Yu, Akash Velu, Eugene Vinitsky, Jiaxuan Gao, Yu Wang, Alexandre Bayen, and Yi Wu. The surprising effectiveness of PPO in cooperative multi-agent games. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022

    2022

  65. [74]

    Linfeng Zhao, Xupeng Zhu, Lingzhi Kong, Robin Walters, and Lawson L.S. Wong. Integrating symmetry into differentiable planning with steerable convolutions. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=n7CPzMPKQl

    2023

  66. [76]

    Sample efficient grasp learning using equivariant models, 2022 b

    Xupeng Zhu, Dian Wang, Ondrej Biza, Guanang Su, Robin Walters, and Robert Platt. Sample efficient grasp learning using equivariant models, 2022 b . URL https://arxiv.org/abs/2202.09468

    2022 arXiv

  67. [77]

    Zinkevich and Tucker R

    Martin A. Zinkevich and Tucker R. Balch. Symmetry in markov decision processes and its implications for single agent and multiagent learning. In International Conference on Machine Learning, 2001. URL https://api.semanticscholar.org/CorpusID:14856766

    2001

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.