MIT develops AI system to improve robot understanding

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory have developed a system that helps robots interpret vague human instructions while using significantly less training data.

The approach, called Masked Inverse Reinforcement Learning (Masked IRL), uses two large language models to clarify tasks and identify the details that matter for safe robot movement.

One model expands ambiguous instructions based on user demonstrations. A second model filters out irrelevant information and highlights factors the robot should include in its motion plan.

The system can help robots understand unstated preferences, such as avoiding a laptop while delivering a coffee mug or keeping a safe distance from a person during a task.

MIT said Masked IRL correctly identified users’ unstated preferences up to 15% more often than comparable methods. Researchers also found that it required nearly five times less demonstration data to learn new tasks.

The approach was tested in simulated environments and on a real robotic arm. The robot completed tasks it had not seen during training, including moving a cup towards a person while avoiding a computer and handing over an object while staying away from nearby obstacles.

Researchers plan to make the system more dynamic by adding cameras, enabling robots to identify relevant objects and ignore distractions in their surroundings visually.

Why does it matter?

Masked IRL could make robots easier to deploy in homes, offices, factories and care environments by reducing the amount of human training needed. The system also addresses a core safety challenge in robotics: people often give vague instructions and leave important preferences unstated. Better interpretation of human intent could help robots work more safely around people, objects and changing environments.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our chatbot!