Robot planning system resolves ambiguous tasks through exploration

Researchers tested a system that lets a robot investigate its surroundings while working out how to complete an underspecified spoken task. In 15 tasks across three indoor and outdoor environments, the system, called CLUE, achieved a success rate within 7 percentage points of a policy given the answers in advance. CLUE proposes possible goals and plans, builds a map as it moves, and revises its plan after interacting with the environment. The researchers report a fourfold success advantage over a language-model planner without that feedback. The tests used a Boston Dynamics Spot and covered object identification, functional reasoning and hidden objects.