acceptodds
Under review as a conference paper at ICLR 2027

Goal-Directed Graph Neural Cellular Automata for Swarm Navigation

Abstract

Cellular automata (CA) are a group of computational models in which each cell repeatedly updates its state through a local rule that depends on its own state and the states of its neighbors. After sufficient iterations, CA are able to exhibit emergent global complexity. Graph cellular automata (GCA) are a generalized version of CA that live and operate on an arbitrary graph rather than a lattice structure. The advantage of GCA is that they allow for dynamic edges between graph nodes instead of fixed distances between cells on a lattice, a feature with strong implications for control-oriented dynamical systems. Prior work used graph neural networks (GNNs) to learn the transition rules of graph cellular automata for flocking imitation. Our central contribution is to extend this framework beyond flocking to goal-directed swarm navigation in both two- and three-dimensional environments. First, we train GNNs to imitate an expert flocking controller that guides swarms from randomized initial configurations through a fixed sequence of waypoints. Then, we train GNNs to imitate the same controller under dynamically updated waypoints sampled throughout the spatial domain. We further evaluate robustness to the removal of up to half of the agents from the middle of the flock during execution, which changes both swarm size and graph topology on the fly. Together, these experiments broaden the role of learned GCA from reproducing dynamics to enabling task-directed collective behavior, highlighting their potential as learned policies for decentralized systems.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Multi-Agent Systems, Decentralized Systems, Graph Cellular Automata, Graph Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ZDWGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper draws on the established link between graph cellular automata and graph neural networks and proposes a graph neural network (GNN) and training procedure to learn a local transition rule for migrating a flock through a waypoint route. The GNN is adapted from [1, Grattarola] to use a mean aggregation to better support continuously changing agent neighborhoods. The paper uses the Boids algorithm, following [1, Grattarola], as an expert controller, but adds a new acceleration term to direct the flock along a series of waypoints. This expert is used in behavior cloning reinforcement learning to train the GNN to learn a local transition that implements the expert's control policy. The loss function has a custom weighting to emphasize how the control for transitioning the flock from moving toward one waypoint to the next. The GNN is evaluated using flocks of 100 agents and a routes of 5 waypoints, in 2 and 3 dimensions. For most of the test routes, the evaluation shows that the GNN successfully visits each waypoint while maintaining a compact, coherent flock, even when a significant number of agents are removed from the center of the flock.

[1, Grattarola] Daniele Grattarola, Lorenzo Livi, and Cesare Alippi. Learning graph cellular automata. In Advances in Neural Information Processing Systems, volume 34, pp. 20983–20994, 2021.

Critical strengths
  1. Strong performance of the method. The GNN controller completes most of the waypoint routes with flock compactness, directional alignment, and speed consistency metrics that meet or improve on the expert controller. The method appears to be an effective solution for the task.
Critical weaknesses
  1. Insufficient comparison with other approaches. The paper justifies comparing only with its expert controller because it is learning a cellular automaton (a local transition rule). There are other proposed methods for learning a local transition rule for flocking, in particular [2, Tolstaya], who learn a decentralized controller trained using [3, Tanner] as the expert (even though the expert is a centralized controller). While the proposed method appears effective, comparing with [2, Tolstaya] where the expert is modified with the same waypoint-seeking acceleration term would help illuminate why the proposed method is the best method.
  2. Dependence on a centralized decision for when a waypoint has been reached by the flock. Although this is a limitation stated in the paper, it means the learned transition is not local as claimed.
  3. Section 3.2 mixes background material with your methodology, and it is difficult to track all the components of your method. I suggest reorganizing this section.

[2, Tolstaya] Ekaterina Tolstaya, Fernando Gama, James Paulos, George Pappas, Vijay Kumar, and Alejandro Ribeiro. Learning decentralized controllers for robot swarms with graph neural networks. In Proceedings of the Conference on Robot Learning, volume 100 of Proceedings of Machine Learning Research, pp. 671–682, 2020.

[3, Tanner] H. G. Tanner, A. Jadbabaie, and G. J. Pappas. Stable flocking of mobile agents part ii: dynamic topology. In Decision and Control, 2003. Proceedings. 42nd IEEE Conference on, volume 2, pages 2016–2021. IEEE, 2003.

Questions for authors
  1. How does the decentralized controller learned in [2, Tolstaya] (where the expert [3, Tanner] has your proposed acceleration term) compare to your proposed method? Or, why is it not possible to compare [2, Tolstaya] for this task?
  2. Instead of using the flock's centroid to determine whether a waypoint has been reached by the flock, how does your method perform when agents decide that a waypoint has been reached using their neighborhood's centroid?
Reviewer_SEP2OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

In this paper, the authors use GNNs to learn behaviors of Graph Cellular Automatons. The main contribution seems to be that the authors extend the work of Gratarola et al. to goal conditioned behavior. They demonstrate the effectivness of the algorithm for a swarm tracking a number of fixed or dynamically updated waypoints. They then test the robustness of the controller to removal of certain swarm members.

Critical strengths

The authors consider an interesting problem in the context of swarm control using learning based approaches.

The authors run a number of extensive experiments to validate the methodology on the general setup.

Critical weaknesses

The extension suggested by the authors seem marginal in novelty.

There is no additional methodological contribution in the paper over Gratarola et al. The authors have merely enlarged the application scope by a little bit.

Questions for authors

The authors should make a clear case in future work, if there is any novel contribution in their work over Gratarola et al. I couldn't see a clear case for this.

Additionally, the authors could consider more general dynamics, such as collision avoidance due to obstacles, multi-agent task allocation. etc. What would be interesting is if for some of these tasks the method of Gratarola et al. doesn't work, and one has to modify the algorithm.

Another interesting question would be how does the method with other permutation invariant architectures.

Reviewer_MBLMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper extends graph neural cellular automata (GNCA; Grattarola et al., 2021) from reproducing Boids flocking behavior to goal-directed swarm navigation in 2D and 3D environments. The expert is a goal-conditioned Boids model in which a fourth steering term pulls each agent’s local-neighborhood centroid toward the active waypoint, preserving the local nature of the expert rule. The GNCA retains the original two-branch architecture but uses mean aggregation instead of sum aggregation. The model is trained using one-step behavior cloning.

Two tasks are evaluated with N = 100 agents:

  1. Fixed waypoints: No goal information is provided as input; instead, the route is encoded in the learned rule. The model achieves completion rates of 40/40 in 2D and 77/80 in 3D. The 3D model requires an octant-based curriculum with experience replay.

  2. Dynamic waypoints: Each agent receives its displacement relative to the active goal, and goals are sampled online. The model achieves completion rates of 40/40 in 2D and 80/80 in 3D. The 3D model requires perturbed expert trajectories; without them, it achieves only 13/80 or 49/80 completions, with most failures resulting from numerical divergence.

A slab-removal test removes 10, 25, or 50 agents during a rollout. The completion rates remain largely unchanged.

Critical strengths

Critical Strengths:

  1. Clear and transparent empirical study. The problem setting is well formalized, and the expert is shown to follow a valid local GCA rule. Complete hyperparameter details and code are provided.

  2. Out-of-distribution robustness evaluation. The agent-removal experiment introduces conditions not encountered during training, changing both the swarm size and graph topology at runtime. The learned rule maintains task completion rates close to those of the expert under these conditions. This provides the most compelling evidence that the learned rule captures the expert’s underlying local mechanism.

Critical weaknesses
  1. Lack of novelty. The paper applies an existing model (GNCA; Grattarola et al., 2021) to a well-known behavior: Boids flocking with an added goal-seeking term. It uses standard behavior cloning and noise injection techniques similar to DART and GNS. The only architectural modification is replacing sum aggregation with mean aggregation. Since the expert itself follows a local rule that the GNCA can represent almost directly, successful imitation is expected. Single-hop GNN controllers trained through imitation learning for decentralized, goal-directed navigation have already been proposed (Tolstaya et al., 2020; Li et al., 2020; Gama et al., 2022). Overall, the paper does not offer sufficient methodological novelty or new insights to meet the expectations of ICLR.

  2. Lack of baselines. The GNCA is compared only against the expert it imitates. Since the expert serves as the training target rather than an independent baseline, the results demonstrate successful imitation but do not establish whether the proposed approach outperforms existing alternatives. Missing comparisons include the original GNCA, E(n)-equivariant GNCA (Gala et al., 2024), GNN-based imitation-learning controllers (Tolstaya et al., 2020; Gama et al., 2022), and a per-agent MLP without message passing. The paper also lacks ablation studies of its own design choices, such as the aggregation function.

Questions for authors

Does the GNCA violate the expert’s separation constraints more frequently than Boids, as measured by the minimum inter-agent distance or the fraction of agent pairs separated by less than rcr_c?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.