SPREAD: Subspace Representation Distillation for Lifelong Imitation Learning

Abstract

A central challenge in lifelong imitation learning (LIL) is enabling agents to acquire new skills from expert demonstrations while retaining knowledge of previously learned tasks. Achieving this requires preserving the low-dimensional manifolds and geometric structures that underlie task representations across sequential learning. However, existing distillation methods, which rely on L2-norm feature matching in the raw feature space, are highly sensitive to noise and high-dimensional variations, often failing to preserve the intrinsic task manifolds. To overcome these limitations, we introduce SPREAD, a geometry-preserving framework that leverages singular value decomposition (SVD) to align the representations of policies from consecutive tasks within low-rank subspaces. This subspace alignment preserves the intrinsic low-dimensional geometry of multimodal features, thereby facilitating stable knowledge transfer, enhancing robustness, and improving generalization across tasks. In addition, we propose a confidence-guided policy distillation strategy that applies a Kullback–Leibler divergence loss restricted to the top-M most confident action samples, emphasizing reliable action modes and improving optimization stability. Empirical results on the LIBERO benchmark demonstrate that SPREAD significantly improves knowledge transfer across tasks, mitigates catastrophic forgetting, and achieves superior overall performance compared to state-of-the-art LIL methods.