Towards zero-forget continual learning for interactive trajectory prediction: a dynamically expandable approach

This paper identifies, analyzes, and addresses case-level forgetting in continual learning for trajectory prediction. We propose the Dynamically Expandable Interactive Trajectory Predictor (DEITP), a novel framework that preserves previously learned knowledge through a dynamic model expansion mechanism. The mechanism regulates expansion timing by assessing model similarity, thereby controlling model growth while preventing catastrophic forgetting. Furthermore, to operate in realistic task-free settings where task identity is unavailable at test time, we introduce a task identification strategy based on a familiarity autoencoder that selects the most appropriate expert for prediction. Extensive experiments on real-world datasets demonstrate that DEITP substantially mitigates forgetting and achieves zero-forgetting performance when task identities are known.