Collaborative Continual Communicative Learning for Black-Box Visual Place Recognition via Place-Aware Distillation

Abstract

Visual Place Recognition (VPR) is a cornerstone of long-term robotic autonomy; however, maintaining performance in non-stationary environments requires robust Continual Learning (CL) to mitigate catastrophic forgetting. While conventional CL paradigms often assume internal access to prior models, the emergence of decentralized multi-robot ecosystems demands a transition toward knowledge transfer between heterogeneous, black-box systems. This study proposes a sophisticated multi-robot CL framework that facilitates autonomous knowledge acquisition from black-box VPR models encountered during exploration. To ensure the framework’s relevance to state-of-the-art systems, we integrate a combinatorial partitioning strategy inspired by CPlaNet, which enables high-resolution spatial discretization while maintaining training data density, and leverage the "place-aware" Stage 1 encoder of Pair-VPR as a robust visual backbone. Central to our methodology is the strategic application of Membership Inference Attacks (MIA) to reconstruct high-fidelity pseudo-training sets, thereby enabling data-free knowledge distillation. To circumvent the sampling inefficiency inherent in high-dimensional latent spaces, we introduce specialized sampling strategies, including Reciprocal Rank and Entropy-based filtering. Extensive experimental validation using the NCLT dataset demonstrates that our approach not only effectively mitigates forgetting but also exhibits remarkable resilience to complex spatial partitions and advanced transformer-based feature representations. These findings confirm the framework’s capacity to facilitate decentralized collaborative learning under strict data-exclusive and black-box constraints in the next generation of autonomous robotic networks.