Masked IRL: how LLMs help robots clarify vague instructions and filter out irrelevant details

Colleagues, I’d like to share an AI update. MIT CSAIL introduced Masked IRL: one LLM clarifies vague commands from demonstrations; a second masks irrelevant details for motion planning.

- Core: one LLM expands incomplete requests using trajectories; the other flags critical scene elements (1/0).
- Impact: nearly 5× fewer demonstrations and up to 15% better recovery of latent preferences; validated in simulation and on a real manipulator.
- Applications: safer object transfer, obstacle avoidance, and honoring user preferences.

Why it matters: reduces human burden and improves robot safety at home and in industry. Which tasks do you think will benefit first?

#AI #robotics #machinelearning #MIT