Teaching robots to perform everyday tasks remains challenging because people rarely provide perfectly clear instructions. Humans naturally leave out details they consider obvious, expecting others to infer their intent from context. Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a new framework called Masked Inverse Reinforcement Learning (Masked IRL) that enables robots to better interpret ambiguous instructions by combining large language models (LLMs) with robot learning algorithms, tells MIT News.
The system uses two complementary LLMs. The first expands a user’s brief instruction by inferring missing details from a small set of human demonstrations. For example, if someone asks a robot to place a coffee mug on a desk without interrupting a video call, the model identifies unstated preferences such as avoiding blocking the computer screen or placing the mug within easy reach. A second LLM then filters the robot’s surroundings, masking irrelevant information while highlighting objects and environmental features that matter for completing the task safely and efficiently.
This focused representation simplifies the robot’s decision-making process and improves the performance of inverse reinforcement learning, a technique that enables machines to infer the goals behind human actions rather than simply copying movements. Because the robot concentrates only on task-relevant information, it requires far fewer demonstrations to learn new behaviors and can adapt more effectively to unfamiliar environments.
In simulations and experiments with a robotic arm, Masked IRL identified users’ unstated preferences up to 15% more accurately than existing methods while using nearly five times fewer demonstrations. The framework also transferred successfully from simulated training environments to real-world robotic tasks without requiring additional retraining, demonstrating its potential for practical deployment.
The research could improve robots used in homes, warehouses, offices, hospitals, and manufacturing facilities, where workers often communicate using natural language rather than rigid commands. Instead of requiring extensive programming for every new task, robots could learn from a handful of demonstrations while understanding the broader intent behind human instructions. By integrating LLM-based reasoning with reinforcement learning, the MIT team offers a promising path toward robots that collaborate more naturally, efficiently, and safely with people in dynamic real-world environments.