Multi-Robot Data-Free Continual Communicative Learning (CCL) from Black-Box Visual Place Recognition Models

TL;DR

This work addresses continual learning for visual place recognition (VPR) in multi-robot systems where target models are private black boxes. It introduces Continual Communicative Learning (CCL), a data-free framework in which a traveler student reconstructs pseudo-training data from black-box teachers via constrained query–response interactions, using Membership Inference Attacks (MIA) reformulated as data-transfer primitives. A prior-based query strategy leverages the student’s VPR priors to focus queries on informative embedding-space regions, reducing knowledge-transfer cost. Experiments on the NCLT multi-session benchmark show substantial gains for low-performing students under modest communication budgets, demonstrating CCL’s potential for scalable, fault-tolerant multi-robot ecosystems.

Abstract

In emerging multi-robot societies, heterogeneous agents must continually extract and integrate local knowledge from one another through communication, even when their internal models are completely opaque. Existing approaches to continual or collaborative learning for visual place recognition (VPR) largely assume white-box access to model parameters or shared training datasets, which is unrealistic when robots encounter unknown peers in the wild. This paper introduces \emph{Continual Communicative Learning (CCL)}, a data-free multi-robot framework in which a traveler robot (student) continually improves its VPR capability by communicating with black-box teacher models via a constrained query--response channel. We repurpose Membership Inference Attacks (MIA), originally developed as privacy attacks on machine learning models, as a constructive communication primitive to reconstruct pseudo-training sets from black-box VPR teachers without accessing their parameters or raw data. To overcome the intrinsic communication bottleneck caused by the low sampling efficiency of black-box MIA, we propose a prior-based query strategy that leverages the student's own VPR prior to focus queries on informative regions of the embedding space, thereby reducing the knowledge transfer (KT) cost. Experimental results on a standard multi-session VPR benchmark demonstrate that the proposed CCL framework yields substantial performance gains for low-performing robots under modest communication budgets, highlighting CCL as a promising building block for scalable and fault-tolerant multi-robot systems.

Figures (4)

• Figure 1: Conceptual illustration of communicative knowledge transfer (KT). Even simple human-to-human communication allows travelers (students) to avoid getting lost by acquiring place-recognition knowledge from locals (teachers) without seeing their internal mental models. In analogy, this study explores a multi-robot Continual Communicative Learning (CCL) framework, where a robot student interacts with black-box teacher VPR models through a query--response protocol and reconstructs pseudo-training data for continual adaptation.
• Figure 2: Examples of input images from independent sessions. Each row presents images from different place classes across four sessions, with each column containing four image samples from the corresponding place class. The grid-based partitioning adopted in this paper is a standard solution to the place class definition issue. However, it results in large intra-class variation, making the CCL-based VPR task more challenging.
• Figure 3: Example of experimental setup. In each scenario, at stage $i = 0$, the student robot trains the VPR model via supervised learning using the data from the place classes it has experienced (blue boxes). In subsequent stages $i = 1, 2, 3$, every time the student encounters a new teacher, it retrains the VPR model via communicative knowledge transfer using the data of the place classes the teacher has experienced (yellow boxes).
• Figure 4: Top-1 accuracy vs. communicative KT cost (the number of pseudo-samples $N$).