Towards Lifelong Aerial Autonomy: Geometric Memory Management for Continual Visual Place Recognition in Dynamic Environments
A new memory framework prevents drones from forgetting old maps, solving a key hurdle for long-term autonomous flight.
A research team led by Xingyu Shao has developed a novel AI architecture to solve a critical problem for long-term drone autonomy: catastrophic forgetting. When drones perform visual place recognition (VPR) over multiple missions in changing environments—like seasonal shifts or new construction—standard AI models forget previously learned geographic features. The team formulates this as a mission-based domain-incremental learning problem and introduces a 'Learn-and-Dispose' pipeline. This framework cleverly decouples geographic memory into two parts: static satellite anchors that preserve global geometric priors, and a dynamic experience replay buffer that retains domain-specific visual features, all while respecting strict onboard storage limits.
To manage the limited buffer, the researchers introduced a spatially-constrained allocation strategy that selects which experiences to keep based on maximizing feature space diversity, not just sample difficulty. This approach proved crucial. In extensive testing on a benchmark of 21 diverse mission sequences, their diversity-driven buffer selection outperformed a random baseline by 7.8% in knowledge retention. The results demonstrate that in unstructured, real-world environments, preserving a wide coverage of structural features is more critical for combating forgetting than methods focused on preserving class averages. This provides a superior balance between plasticity (learning new things) and stability (remembering old things), ensuring drones remain robustly localized regardless of the order of their missions.
- Proposes a 'Learn-and-Dispose' memory framework that splits drone mapping knowledge into static geometric anchors and a dynamic feature buffer.
- Diversity-driven buffer selection strategy outperforms random baseline by 7.8% in knowledge retention across 21 test mission sequences.
- Solves catastrophic forgetting for drones, enabling lifelong adaptation to environmental changes like seasons or new construction without losing old map data.
Why It Matters
Enables drones and robots to operate autonomously for months or years, reliably navigating through cities as they evolve, which is vital for delivery, inspection, and surveillance.