Vision‑Language‑Action (VLA) policies use pretrained vision‑language models (VLMs) to guide robot action generation. VLMs produce hierarchical visual‑semantic features that evolve from low‑level geometry to high‑level language‑aligned semantics, so different manipulation tasks may require different mixtures of these layer representations. At the same time, the action module creates intermediate states during computation, which can be useful for later decisions. Existing VLA interfaces expose VLM information through fixed layer assignments for each action layer, and intermediate action states are only propagated implicitly via residual streams, preventing explicit reuse. To address this, we introduce LayerRoute, an action‑conditioned representation routing interface that enables adaptive access to both VLM layers and action representations.
LayerRoute consists of two mechanisms:
- Layer Mixture Router dynamically forms mixtures of cached VLM representations, weighting layers as needed;
- Action‑State Reread allows later action layers to reread and reuse earlier action states. Across diverse simulation and real‑world benchmarks, LayerRoute consistently improves StarVLA‑$\pi$ and $\pi_{0.5}$, achieving up to $7.2$ gains on LIBERO Long while adding only $0.31\%$ (relative) or $3.87\%$ (absolute) extra parameters. Ablation studies confirm the benefit of action‑conditioned layer routing, and routing analyses reveal structured allocation patterns across action layers and task settings.
Review: LayerRoute breaks the rigidity of traditional VLA interfaces by explicitly mixing layers and reusing action states, delivering notable performance gains with minimal parameter overhead and offering a more flexible feature utilization framework for vision‑language‑driven robot control.