ChatVLA presents a breakthrough approach to robot intelligence by integrating vision, language, and action capabilities into a unified model.

• Solves key challenges in robot training: spurious forgetting and task interference
• Achieves balanced performance across vision-language understanding and robot control
• Demonstrates success on 25 real-world manipulation tasks
• Introduces a novel training paradigm that preserves multimodal alignment

This research advances engineering by creating robots that can simultaneously perceive, understand, and interact with their environment in a human-like manner, potentially transforming automation capabilities across industries.

ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model