Teaching a robot to understand what you really want — not just what you say — has long been one of the harder problems in robotics. A human might ask a robot to “bring coffee without disturbing the meeting,” but that leaves out crucial details: stay away from the laptop, don’t block the camera, approach from the side. Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a method called Masked Inverse Reinforcement Learning (Masked IRL) that lets robots infer those unspoken preferences automatically, using nearly five times fewer demonstration examples than conventional approaches.
What Is Masked Inverse Reinforcement Learning?
Masked IRL is a framework that combines inverse reinforcement learning with large language models to teach robots manipulation tasks from minimal human input. Instead of requiring dozens of meticulously recorded demonstrations or pages of written instructions, the system lets a human physically guide a robot through a task a few times — a technique called kinesthetic demonstration — and then uses LLMs to figure out which details matter and which can be ignored. “Our approach could come in handy when a human interacts with a robot but doesn’t want to spell out all the details of a task,” says Minyoung Hwang, an MIT PhD student and CSAIL researcher who is lead author on the paper presenting the project. “We’re minimizing human effort by enabling machines to get to the bottom of what users really want.”
How Masked IRL Works: Two LLMs in Tandem
The system operates in two stages. First, an LLM compares the motion trajectory from a human demonstration — the specific sequence of positions the robot was guided through — to the shortest possible path for completing the action. The same LLM also elaborates on ambiguous instructions, turning a vague request like “stay close” into something more precise like “stay close to the surface of the table.” This clarification step helps the model understand why the demonstrated motions matter.
A second LLM then evaluates the environment, examining elements such as obstacle positions and the shape of the target object. It assigns each element a score: “1” for important, “0” for irrelevant. For example, whether a human leaned on a table during the demonstration is scored as 0 and discarded. Details marked as 1 are incorporated into the final motion plan. This masking process is what gives the method its name and its efficiency advantage.
Performance Gains and Real-World Testing
In both simulation experiments and tests with a real robotic arm, Masked IRL outperformed comparable baselines. The system correctly identified users’ unstated preferences up to 15 percent more often than alternative approaches. It also required fewer demonstrations to learn how to move objects like a coffee mug around obstacles such as a laptop on a table.
After training on 50 kinesthetic demonstrations, a real robot arm successfully performed tasks it had not seen during training: moving a cup toward a human while avoiding a computer, wiping a table while staying close to the surface, and handing a bag of chips while staying away from both the person and the table. The robots performed better when the LLM clarified instructions than when they attempted to follow vague requests directly.
Next Steps: Adding Visual Understanding
The current system relies on sensor data and motion logging, but the team plans to make Masked IRL more dynamic by adding camera input. This would allow the robot to capture images of its surroundings and visually highlight relevant elements while ignoring others — for instance, spotting bananas near a toy and disregarding them before picking up the target object. The research was supported in part by the Tata Group through the MIT Generative AI Impact Consortium Award and the Department of Defense. Hwang co-authored the paper with CSAIL colleagues Alexandra Forsey-Smerek, Nathaniel Dennler, and MIT Assistant Professor Andreea Bobu. The team will present the work at the 2026 IEEE International Conference on Robotics and Automation in June.
What This Means for Developers and Roboticists
Masked IRL addresses a practical bottleneck in deploying robots in human environments: the cost of collecting demonstration data. By requiring roughly five times fewer examples and automatically resolving ambiguity in instructions, the approach makes it more feasible to train robots for tasks in homes, offices, and factories where explicit programming is impractical. For engineers working on robotic manipulation, the paper demonstrates that combining inverse reinforcement learning with LLM-based clarification and masking can yield significant efficiency gains without sacrificing task accuracy. The full paper is available on arXiv under identifier 2511.14565 for those who want to study the technical details and implement the approach.
