SPREAD: Subspace Representation Distillation for Lifelong Imitation Learning

A team of researchers including Kaushik Roy and Giovanni D'urso has published a paper introducing SPREAD (Subspace Representation Distillation for Lifelong Imitation Learning), a novel framework designed to tackle a fundamental problem in AI robotics: catastrophic forgetting. The core challenge in lifelong imitation learning (LIL) is enabling agents, like robots, to sequentially acquire new skills from expert demonstrations while perfectly retaining all prior knowledge. Existing methods, which often rely on simple L2-norm feature matching, are sensitive to noise and fail to preserve the intrinsic low-dimensional geometric structures, or 'manifolds,' that define different tasks.

SPREAD's innovation lies in its geometry-preserving approach. Instead of working in the raw, high-dimensional feature space, it employs singular value decomposition (SVD) to project and align policy representations into stable, low-rank subspaces. This maintains the underlying geometric relationships between multimodal features, leading to more robust transfer and generalization. The team also developed a complementary confidence-guided distillation strategy, which applies a Kullback-Leibler divergence loss only to the top-M most confident action samples, focusing learning on reliable data and improving optimization stability.

Experiments conducted on the standard LIBERO lifelong imitation learning benchmark demonstrate that SPREAD delivers a substantial performance leap. The framework shows marked improvements in knowledge transfer between tasks and effectively mitigates the forgetting of previously learned skills, achieving state-of-the-art results. This work, accepted for IEEE ICRA 2026, represents a significant step toward creating more adaptable and capable robotic systems that can learn continuously in real-world environments.