HAMSTER introduces a hierarchical vision-language-action model that enables robots to generalize skills in open-world environments despite limited training data.

• Leverages cheaper off-domain data (videos, sketches, simulations) to overcome robotics data scarcity
• Employs a hierarchical approach that connects high-level reasoning with low-level execution
• Demonstrates improved generalization capabilities for complex manipulation tasks
• Addresses fundamental challenges in robotics by applying foundation model principles

This research represents a significant advance in robotics engineering by creating more adaptable manipulation systems that can operate in unstructured environments with less specialized training data—potentially revolutionizing industrial automation and human-robot collaboration scenarios.

HAMSTER: Hierarchical Action Models For Open-World Robot Manipulation