Fed-EC: Bandwidth-Efficient Clustering-Based Federated Learning For Autonomous Visual Robot Navigation

TL;DR

Non-IID data across robots in outdoor navigation hinders single global FL models. Fed-EC introduces embedding-based clustering with DBSCAN to create cluster-specific models, using mean embeddings $V_r = \sum_{i=1}^{|D_r|} v_i$ to capture data distribution and clustering by distance $||V_r - V_j||$ to group similar robots, which reduces communication by sharing embeddings and local weights. Experiments show Fed-EC matches centralized learning performance with 23x per-robot bandwidth savings and outperforms purely local training, with the added benefit of transferring cluster models to new joiners. This approach enables scalable, privacy-preserving, bandwidth-efficient autonomous navigation in diverse environments with continual adaptation and transferability.

Abstract

Centralized learning requires data to be aggregated at a central server, which poses significant challenges in terms of data privacy and bandwidth consumption. Federated learning presents a compelling alternative, however, vanilla federated learning methods deployed in robotics aim to learn a single global model across robots that works ideally for all. But in practice one model may not be well suited for robots deployed in various environments. This paper proposes Federated-EmbedCluster (Fed-EC), a clustering-based federated learning framework that is deployed with vision based autonomous robot navigation in diverse outdoor environments. The framework addresses the key federated learning challenge of deteriorating model performance of a single global model due to the presence of non-IID data across real-world robots. Extensive real-world experiments validate that Fed-EC reduces the communication size by 23x for each robot while matching the performance of centralized learning for goal-oriented navigation and outperforms local learning. Fed-EC can transfer previously learnt models to new robots that join the cluster.

Figures (9)

• Figure 1: Workflow of Fed-EC. (a) Participating robots learn and communicates a mean embedding vector and local model weight to the server. The server clusters the robots using the mean embedding vector and aggregates local models in each cluster to learn a model which is shared with the robots based on their cluster identity. (b) The robots navigate to the a given GPS goal using the learnt model which takes as input RGB and depth images from the front facing camera. (c) If a new robot is deployed it computes a mean embedding and shares it with the server. The server assigns a cluster to the robot and sends the respective cluster model to the robot to use.
• Figure 2: Terrasentia Robot with ZED2 camera and GNSS used as a testbed for our experiments.
• Figure 3: Sample Data and Traversability Labels collected across different terrains.
• Figure 4: Common obstacles encountered by the robot. Left to right: wall, fire hydrant,light pole, tree.
• Figure 5: Effect of increasing computation per robot.