MoLe-VLA introduces an innovative architecture that enables efficient robot manipulation by dynamically skipping unnecessary computation in multimodal large language models.

• Achieves 40.1% reduction in computational requirements without significant performance loss
• Employs a mixture-of-layers approach that intelligently determines which model layers to use for different inputs
• Demonstrates real-world viability in robot manipulation tasks with comparable accuracy to full models
• Provides a path to deployment for complex vision-language models in resource-constrained robotic systems

This research addresses a critical engineering challenge in robotics: enabling sophisticated language-vision models to run efficiently on robots with limited computational resources, significantly advancing practical factory automation applications.

MoLe-VLA: Dynamic Layer-skipping Vision Language Action Model via Mixture-of-Layers for Efficient Robot Manipulation