LLaRA transforms robot action policy into visuo-textual conversations, enabling efficient transfer of pretrained Vision Language Models (VLMs) to robotics with minimal demonstrations.

• Formulates robot control as a conversation between visual inputs and textual commands
• Enables more efficient learning from limited robot demonstration data
• Bridges the gap between powerful VLMs and practical robotic applications
• Demonstrates effectiveness in both simulated and real-world tasks

This research advances robot engineering by tackling a fundamental challenge: how to leverage powerful language models for physical control systems without requiring massive amounts of specialized robotics data.

LLaRA: Supercharging Robot Learning Data for Vision-Language Policy