HyperDCM helps robots remember navigation across scenes using hyperbolic memory
Robots no longer forget how to navigate old environments when learning new ones.
Continual learning in robotics has a persistent problem: as a robot learns to navigate a new environment, it often forgets how to navigate previously learned ones — a phenomenon known as catastrophic forgetting. To address this, a research team introduces HyperDCM (Hyperbolic Dynamic Cluster Memory), a structure-aware memory mechanism designed specifically for diffusion policy-based visual navigation.
HyperDCM works in three stages. First, it extracts semantic scene triples (e.g., object-relation-object) from RGB observations using large vision-language models. These triples are then encoded into scene graph embeddings via a Relational Graph Convolutional Network (R-GCN). Critically, the embeddings are projected into hyperbolic space, which better captures hierarchical and structural information than Euclidean space. Finally, a dynamic clustering and structure-sensitive update strategy selects the most representative samples for memory replay, preserving knowledge diversity. On multi-scene indoor and outdoor datasets, HyperDCM outperformed several continual learning baselines adapted to diffusion policy navigation, showing strong retention of past navigation capabilities and improved generalization to new scenes.
- Uses scene graph embeddings from large vision-language models to capture object-relation structures in navigation environments.
- Projects embeddings into hyperbolic space to improve structural separability and long-term retention.
- Dynamic clustering selects representative memories for replay, reducing catastrophic forgetting across scenes.
Why It Matters
Enables robots to continuously learn new environments without forgetting old ones, critical for practical deployment in homes and warehouses.