A new robot-planning framework is built to say no when instructions do not add up.
Researchers behind a framework called Risk-Aware Semantic Grounding argue that large language models acting as robot planners are too eager to comply, even when an instruction is vague, contradicts the physical environment, or conflicts with itself. Instead of letting the LLM jump straight to generating a plan, the system first scores the instruction for three kinds of risk: ambiguity, hallucination, and semantic conflict. Based on that score, the robot can choose to execute the command, ask for clarification, or refuse it outright. The team also built a benchmark called TRUST-NAV that mixes ordinary navigation tasks with instructions deliberately designed to trip up a planner.
Most robot-planning research still measures success by whether the task gets done, which rewards models for guessing confidently even when they should not. This shifts the yardstick toward knowing when not to act, a distinction that matters more once a robot is moving through a room with people in it rather than just outputting text. The researchers report that conventional LLM planners handle valid navigation tasks well but are comparatively worse at catching ambiguous or conflicting instructions than the risk-aware approach, though the abstract does not disclose the specific numbers behind that gap.
Call it a robot learning to ask what do you mean before it starts pushing furniture.