Stanford's LIRA boosts robot success rates by 18.9%
New VLA model LIRA uses cross-layer routing to slash robot task failures...
Researchers from Stanford University and collaborators have developed LIRA (Local Cross-Layer Information Routing), a novel mechanism designed to bridge the gap between vision-language models (VLMs) and robot action decoders in Vision-Language-Action (VLA) systems. Unlike existing approaches that either expose limited representation hierarchies or rigidly align decoder blocks to specific VLM layers, LIRA introduces a dynamic, depth-aware routing system. It leverages task-token features and intermediate VLM states to create LIRA Query features, which are then aggregated by Parallel Fusion Blocks in a localized, depth-aligned manner.
The method’s key innovation lies in its ability to improve task-token feature conditioning without altering the underlying VLM backbone, action decoder, or training pipeline. In empirical tests across benchmarks like LIBERO, LIBERO-Plus, and CALVIN ABC→D, LIRA demonstrated significant gains over the baseline VLA-Adapter (0.5B parameters). Notably, in zero-shot transfer to LIBERO-Plus, LIRA achieved an 18.9-point improvement in average success rate (59.1% → 78.0%), highlighting its robustness to distribution shifts. The team plans to release code and model checkpoints upon paper acceptance.
- LIRA improves robot task success rates by 18.9% (59.1% → 78.0%) in zero-shot tests using a 0.5B-parameter VLA model.
- The method dynamically routes intermediate VLM features across layers via depth-aligned Parallel Fusion Blocks, enhancing task-token conditioning.
- LIRA maintains compatibility with existing VLA architectures and training recipes, requiring no modifications to the backbone or decoder.
Why It Matters
LIRA’s cross-layer routing could accelerate real-world robot deployment by reducing task failures in novel environments.