MIT CSAIL Unveils Masked IRL, Cutting Robot Training Data by Nearly 5 Times
Updated
Updated · MIT News · Jun 26
MIT CSAIL Unveils Masked IRL, Cutting Robot Training Data by Nearly 5 Times
1 articles · Updated · MIT News · Jun 26
Summary
MIT CSAIL said its new Masked IRL system lets robots learn from vague human instructions with nearly five times less demonstration data, reducing the need for extensive "show and tell" training.
Two large language models drive the method: one expands ambiguous prompts using a user’s demonstrated motion, and another filters the environment to keep only task-relevant details for the robot’s motion plan.
In tests, the system inferred unstated user preferences up to 15% more often than comparable baselines, helping robots route objects around obstacles such as moving a mug around a laptop.
After 50 kinesthetic demonstrations, a real robotic arm completed unseen tasks including handing over a cup while avoiding a computer, wiping a table while staying close, and delivering chips while keeping distance from a person and table.
The researchers plan to add cameras so robots can visually identify and ignore irrelevant nearby objects, extending the approach for homes, offices and factories.
As AI data costs soar, can MIT's 'masking' robot brain solve the industry's billion-dollar training problem?
MIT's AI learns with 80% less data. Is this the breakthrough that finally puts intelligent robots into our homes?
If robots now 'guess' our intentions, what new safety risks emerge when they inevitably guess wrong?
Masked IRL Sets New Standard: 92% Success Rate and 30% Faster Training for Robots Learning from Language and Demonstrations
Overview
Masked IRL (M-IRL) is a novel framework for robot learning that has been accepted for presentation at ICRA 2026. By combining the strengths of human demonstrations and natural language instructions, Masked IRL makes the robot learning process more intuitive, efficient, and robust. Traditionally, robots learned either from demonstrations, which show 'how-to' steps, or from language, which provides goals, but each method had limitations. Masked IRL overcomes these by using large language models to blend both approaches, allowing robots to better understand tasks and adapt to new situations, marking a significant advancement in the field.