Separating Routing Organization from Task Success: Matched Evidence from Prototype-Routed MoE Policies

Mixture-of-experts (MoE) policies can represent the different control regimes within a manipulation task. However, their routing partitions are usually learned without using the kinematic structure from the demonstrations. Therefore, we study whether trajectory-derived kinematic regimes can serve as an initialisation prior, obtained by segmenting and clustering the demonstrations. Stage 1 is trained with action prediction and rule-derived phase supervision. Regime-conditioned centroids of the Stage 1 latent representation then initialise hard-routing prototypes before joint fine-tuning. A matched prototype-routing ablation compares trajectory-derived, random, and rule-based prototype initialisation. Phase-expert alignment (NMI) and routing-switch rate are measured on validation demonstrations, separately from closed-loop rollout success. A learned softmax gate serves as an architectural diagnostic. Trajectory-derived initialisation yields higher mean final NMI (0.67 on Can; 0.51 on Square) than random and rule-based initialisation (0.07–0.09), with lower mean routing-switch rates (0.04–0.07 versus 0.10–0.11). Pooled rollout success is 76.0% on Can, compared with 73.3% and 68.7% for random and rule-based initialisation, and 26.7% on Square, compared with 28.0% and 31.3%. Under matched conditions, trajectory-derived initialisation produces distinct offline routing partitions, but closed-loop success differences across three independently trained seeds are not statistically resolved after Holm adjustment.