Event

HIDRA: A Routing Framework for Replay-Free Lifelong Imitation Learning

1 reports · 1 sources

Fanqi Yu and co-authors propose HIDRA, a hierarchical dual-attention routing framework for replay-free, single-pass lifelong imitation learning. It decouples instruction-level expert retrieval from context-dependent refinement and introduces key-level geometric regularization that constrains intra-task alignment and inter-task separation, in order to mitigate routing interference under representation drift. The authors state that under replay-free, single-pass settings, sequential updates progressively distort the representation space, making expert-based methods particularly sensitive to routing interference, with expert selection degrading over time. The authors evaluate HIDRA on multiple LIBERO benchmark suites, including harder variants such as heterogeneous task sequences and paraphrased language instructions; results show that HIDRA outperforms replay-free and replay baselines on AUC and forward transfer while maintaining low forgetting. This work is an author preprint; the authors state it has been accepted to NeurIPS 2026, and the code has been made public.