LIRA: Local Cross-Layer Information Routing for Vision-Language-Action Decoding

2026-08-06

AI Summary

Existing vision-language-action (VLA) models often employ rigid one-to-one mappings or utilize only a limited set of intermediate features from vision-language models (VLMs), hindering effective integration of complementary cross-depth information. This work proposes Local Inter-layer Routing with Attention (LIRA), a mechanism that dynamically assigns, for each parallel fusion block, a depth-aligned local window centered on the corresponding VLM layer—without modifying the backbone, decoder, or training pipeline. LIRA aggregates neighboring query features and fuses them with task tokens and proprioceptive inputs to predict actions. Introducing depth-aware local inter-layer routing for the first time, LIRA significantly outperforms the VLA-Adapter baseline across LIBERO, LIBERO-Plus, CALVIN ABC→D, and real-world robotic tasks. Notably, in zero-shot transfer to LIBERO-Plus, it improves average success rate from 59.1% to 78.0%, a gain of 18.9 percentage points.

Abstract

Vision-Language-Action (VLA) models transform representations from pretrained vision-language models (VLMs) into robot actions, yet the interface that routes intermediate VLM features into action decoders remains underexplored. Existing designs either expose only a narrow part of the representation hierarchy or rigidly match each decoder block to one VLM layer, restricting access to complementary task evidence across depths. We introduce LIRA, a local cross-layer action-conditioning mechanism that formulates VLM-to-action conditioning as depth-aware information routing. LIRA operates on task-token features and LIRA Query features derived from intermediate VLM states, then assigns each Parallel Fusion Block a depth-aligned local window centered on its corresponding VLM layer. Parallel Fusion Blocks aggregate neighboring LIRA Query features and integrate them with task-token features and proprioceptive inputs before action prediction. This routing interface leaves the backbone architecture, action decoder, and supervised training recipe unchanged. Across LIBERO, LIBERO-Plus, CALVIN ABC$\rightarrow$D, and real-world manipulation, LIRA improves the principal aggregate metrics over the VLA-Adapter baseline under the same 0.5B-parameter configuration. In zero-shot transfer to LIBERO-Plus, LIRA increases average success from 59.1% to 78.0%, an 18.9-point gain indicating improved robustness under controlled distribution shifts.

Problem

• Vision-Language-Action
• cross-layer routing
• action decoding
• representation hierarchy
• information routing

Innovation

• LIRA
• Vision-Language-Action
• cross-layer routing
• depth-aware fusion
• zero-shot transfer