Distill: New AI method decodes true intent in human-robot communication

As robots enter homes and workplaces, users rely on natural language or end-user programming to specify tasks. Yet these modalities are fundamentally flawed: natural language is imprecise and ambiguous, while programming is overly rigid. A person might say “clean the dishes and then put them away” but actually mean “handle the dishes in any order that makes sense.” The disconnect between what users say and what they intend remains a core challenge for human-robot interaction.

To solve this, Ting Li and David Porfirio propose Distill, a system that automatically refines initial task specifications into a more accurate representation of user intent. Distill applies three transformations: removing unnecessary steps (e.g., skipping a redundant “open dishwasher” if the robot already knows), generalizing the meaning behind individual steps (e.g., “wash” becomes “clean”), and relaxing ordering constraints (e.g., allowing parallel execution). In a crowdsourced study using a web interface, Distill consistently produced specifications that better matched users' true goals. The work, published on arXiv, offers a practical pathway for making robot instructions more robust without requiring users to learn new programming languages.