📄 中文摘要
终身模仿学习（LIL）中的一个关键挑战是使智能体能够从专家示范中获取新技能，同时保留先前的知识。这需要在连续学习中保持任务表示的低维流形和几何结构。现有的蒸馏方法依赖于在原始特征空间中进行 L2 范数特征匹配，容易受到噪声和高维变异的影响，往往无法保留内在的任务流形。为了解决这一问题，提出了一种名为 SPREAD 的几何保留框架，该框架采用奇异值分解（SVD）在低秩子空间中对任务间的策略表示进行对齐。这种对齐方式保持了多模态特征的底层几何结构。
SPREAD：用于终身模仿学习的子空间表示蒸馏
📄 English Summary
SPREAD: Subspace Representation Distillation for Lifelong Imitation Learning
A key challenge in lifelong imitation learning (LIL) is enabling agents to acquire new skills from expert demonstrations while retaining prior knowledge. This necessitates the preservation of low-dimensional manifolds and geometric structures underlying task representations across sequential learning. Existing distillation methods, which rely on L2-norm feature matching in raw feature space, are sensitive to noise and high-dimensional variability, often failing to preserve intrinsic task manifolds. To address this, a geometry-preserving framework named SPREAD is introduced, employing singular value decomposition (SVD) to align policy representations across tasks within low-rank subspaces. This alignment maintains the underlying geometry of multimodal features.