This research evaluates and addresses the visual generalization capabilities of modern Vision Language Action (VLA) robotic foundation models, improving their ability to function in diverse visual environments.

• Identifies significant performance drops when robotic models encounter visual domains different from their training data
• Proposes a new evaluation benchmark to systematically test visual domain adaptation in robotics
• Demonstrates that fine-tuning with carefully selected visual data can significantly improve cross-domain performance
• Shows that improved visual adaptation leads to better real-world task execution in robotics

This engineering advancement is crucial for developing robotic systems that can reliably operate across varied real-world environments without requiring domain-specific retraining.

ReVLA: Reverting Visual Domain Limitation of Robotic Foundation Models