acceptodds
Withdrawn from ICLR 2027

ATLAS-HFL: Learning-Aware Anchor Graphs for Closed-Loop Federated Learning in Dynamic UAV Swarms

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Dynamic UAV swarms must support federated task learning under changing topology, limited resources, heterogeneous data, and intermittent cluster-head failures. Existing UAV clustering methods usually treat the cluster structure as a communication route or aggregation scaffold. However, in dynamic UAV federated learning, a useful organization structure should encode not only physical feasibility and resource sustainability, but also task relations that guide future model cooperation. The observations suggest that swarm organization should be maintained as an evolving learning-aware state rather than a one-time clustering result. Following this reasoning, a hierarchical federated learning framework called ATLAS-HFL is developed through a sparse anchor graph. ATLAS-HFL represents UAVs with physical states, resource states, and compact local learning states without constructing a full pairwise UAV graph. The learned UAV-to-anchor associations determine dynamic clusters, support primary-backup cluster-head election, and produce task-aware inter-cluster cooperation weights for cluster-specific initialization in the next control period. Updated local training states then refine the anchor graph, forming a closed loop between federated learning and swarm organization. Experiments across heterogeneous federated learning settings, cluster-head failure scenarios, scalability tests, and network simulations show that ATLAS-HFL improves task performance, preserves service availability under cluster-head failures, and enhances communication reliability while maintaining comparable network lifetime.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Dynamic UAV Swarms, Learning-Aware Anchor Graphs, Primary-Backup Cluster Heads, Closed-Loop Federated Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_EiopOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ATLAS-HFL, a hierarchical federated learning framework for dynamic UAV swarms. It combines physical, resource, and local learning states through a sparse multi-view anchor graph. The learned graph determines dynamic clusters, supports primary-backup cluster-head (CH) election, and guides task-aware inter-cluster model cooperation for the next control period. Experiments evaluate classification performance, statistical heterogeneity, CH failures, scalability, and NS-3 network performance.

Critical strengths

S1. The framework addresses data heterogeneity alongside mobility, communication, and resource constraints. Using the learned organization to guide subsequent model initialization provides a coherent connection between swarm organization and task learning.

S2. The anchor representation avoids constructing a full pairwise UAV graph. This offers a compact way to maintain collaboration structure.

S3. The evaluation covers both learning and system performance.

Critical weaknesses

W1. The accuracy advantage is limited. In Table 1, Ditto outperforms ATLAS-HFL on three of the five datasets, while ATLAS-HFL only achieves the highest mean accuracy on FEMNIST and DomainNet.

W2. The paper leaves important experimental details unspecified. These include model architectures and parameter counts, local training settings, key hyperparameters, the construction and dimensions of the learning summaries in Eq. (5), and the NS-3 configuration. The evaluation protocol should also clarify whether accuracy is measured using global or cluster-specific models and how results are aggregated across UAV test sets.

W3. The comparison with closely related methods is incomplete. X-CFL is cited but not evaluated. Related methods such as pFedGraph, FedAMP, and FedFomo infer collaboration relationships or construct personalized model combinations, yet are neither discussed nor compared.

W4. The datasets introduced at line 291 are listed without source citations, and the corresponding dataset references are absent from the bibliography.

Questions for authors

Please clarify the experimental protocol, provide comparisons with closely related methods, and quantify the additional computational and communication costs.

Reviewer_3417OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper develops a federated learning solutions for clustering UAV swarms. The key contributions lies in the consideration of system dynamics and evolution of cooperation in the swarm. The works also accounts for physical resources.

Critical strengths
  • The study of federated learning in dynamic, real-world settings is a meaningful problem.
  • The experimental results on UAV population are interesting.
Critical weaknesses
  • The overall formulation of the problem does not have any unique features captured that are specific to UAV swarms. The problem statement and solution seem like a generic clustering problem.
  • The problem of clustering has been studied from many perspectives in the past, including using stochastic game theory to capture dynamics. It is not clear how this work goes beyond that existing body of work. -In terms of federated learning, the paper simply applies known algorithms, and does not provide advances to the learning techniques and/or their applications (e.g., federated learning has been extensively applied already to drone swarms).
  • The experimental results show minimal gains from the proposed approach compared to more standard learning algorithms.
  • It is not very clear what networking assumptions are made to capture the communication angle of the work.
Questions for authors
  • What are the unique features of UAVs that motivated a new federated learning approach?
  • What are the key contributions to the learning algorithm and how does that compare to prior works?
  • What is the communication system model and how does it expand beyond existing UAV swarm communication works?
  • How do you justify the added complexity of the solution vs the relatively small gains?
  • Can you clarify whether you can handle fast dynamics and at which timescale?
  • What happens if the dynamics change before the learning algorithm has converged?
Reviewer_TWe1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies hierarchical FL in dynamic UAV swarms and proposes ATLAS-HFL, which combines physical, resource, and learning information to construct the graph. Then the graph is used for clustering, primary CH selection, and model cooperation. Experiments cover five image datasets and NS-3 network simulation. The problem is practical, but the advantage over existing joint clustering and graph-based FL methods is not clear. The experiments do not clearly show how much each part of the method helps. More comparisons with existing methods, together with computation and communication measurements on real devices are needed to support the conclusions.

Critical strengths

S1.The paper addresses a practical problem: supporting federated learning in dynamic UAV swarms with limited resources and changing network conditions.

S2.The method considers physical, resource, and learning information together for guiding clustering, CH selection, and model cooperation.

S3.The evaluation covers five image datasets and NS-3 simulations, examining both learning performance and network behavior.

Critical weaknesses

W1. This paper already acknowledges that X-CFL combines learning information with physical and resource states. Therefore, authods should explain what the anchor graph adds beyond these ideas and compare against closer baselines. Authods should add comparison of replacing the anchor graph with a simpler clustering method.

W2. The paper includes NS-3 results, but it does not explain how packet loss, delay, and energy depletion affect model transmission and aggregation. It is also unclear how updated learning states change the same simulated network. A physical UAV testbed will strengthen the evaluation.

W3. The main comparison includes CIFAR-10, CIFAR-100, FEMNIST, and DomainNet. However, the non-IID sensitivity analysis is limited to Fashion-MNIST. It is a very easy dataset. Please repeat these experiments on CIFAR or DomainNet to show whether the reported robustness holds on more demanding tasks.

W4. The full method mixes a global model with task-weighted cluster models, while the ablated version uses only its own cluster model. This removes both global mixing and cross-cluster cooperation. Please keep the global mixing unchanged. The current results, do not clearly establish the benefit of the proposed method.

W5. Figure 1 is difficult to read at normal page scale. Several labels and the embedded performance plot are too small.

W6. Including physical and resource states in the clustering features does not guarantee that every member can communicate with its selected CH. Equation (20) only constrains cooperation between CHs. Please explain how infeasible clusters, empty clusters, and clusters with few members for a backup CH are handled.

Questions for authors

Q1. Please explain how infeasible clusters, empty clusters, and clusters with few members for a backup CH are handled.

Q2. Please what the anchor graph adds beyond these ideas and compare against closer baselines. Authods should add comparison of replacing the anchor graph with a simpler clustering method.

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

Related papers

Discussion (0)

Sign in to comment.