CoRDE: Concept-Prior Routed Diffusion Experts for Structural Generalization in Robot Manipulation

Diffusion models excel at capturing multi-modal action distributions in robot imitation learning. However, in multi-task and long-horizon scenarios, monolithic architectures lack structural generalization capabilities, suffering from gradient conflicts between distinct semantic sub-stages. While pure data-driven Mixture-of-Experts (MoE) methods introduce labor division, they frequently trigger routing collapse, and instantiating full-scale experts causes parameter explosion and high expansion costs. To address these issues, we propose Concept-prior Routed Diffusion Experts (CoRDE), a structure-guided variational distillation framework. CoRDE extracts semantic distributions from a frozen concept encoder to guide the variational posterior responsibility via a learnable soft mapping matrix. This mechanism enforces a dual-entropy dynamics conservation: it minimizes routing entropy to guarantee macroscopic cognitive certainty, while preserving the full-rank variance of the stochastic diffusion term to maintain behavioral diversity. To overcome parameter inflation, CoRDE employs a parameter-efficient expert pool using Low-Rank Adaptation (LoRA) on a shared frozen backbone. Rigorous mathematical proofs demonstrate that the mixture score field of low-rank experts strictly approximates the teacher distribution in the least-squares sense, avoiding rank-deficiency-induced diversity loss and ensuring high-fidelity generation. Empirical evaluations confirm that, compared to existing baselines, CoRDE systematically reduces routing collapse, forming robust, semantically aligned expert allocations while achieving superior action quality and incremental learning efficiency.