Robot Navigation: Contrastive Learning of DistilBERT-Based Semantic Embeddings for Scene Understanding

Conventional robot navigation techniques in dynamic indoor environments frequently rely almost exclusively on geometric sensor data and object detection labels, ignoring important factors needed for context-aware, flexible decision-making. Embedding-based methods can capture nuanced domain-relevant properties, which justifies using a pre-trained language model to incorporate context-sensitive attributes into a robot navigation system by leveraging detected object labels. This paper introduces Supervision Signals for Semantic Embeddings (3SE), a model designed for fine-tuning DistilBERT embeddings using high-level context mapping, parent synonyms, and attribute prototypes with a self-supervised contrastive learning approach and custom attribute propagation losses. Evaluations with UMAP visualisations and quantitative similarity metrics reveal that the revised embeddings isolate domain-critical ideas while preserving semantic meaning. Results indicate that 3SE has the potential to support a semantic robot navigation system's cognitive layer by learning intrinsic attributes and associated hazards of objects through text labels and producing cohesive embeddings. This framework is intended to help autonomous mobile robots make more context-aware navigation decisions in a real-world dynamic indoor environment via enhanced semantic scene understanding in the future, and it can be applied to other domain-specific applications. (Abstract)