KnowNo

Has an LLM planner ask a person when uncertain, using conformal prediction to give a statistical success guarantee.

KnowNo was released in July 2023 by Princeton University and Google DeepMind, and won the CoRL 2023 best student paper award. When a large language model plans robot tasks, an ambiguous instruction often leads it to confidently produce the wrong plan. KnowNo turns each planning step into a multiple-choice question: it first has the model list several candidate actions, then uses conformal prediction (a method that needs only a small amount of calibration data and gives a statistical coverage guarantee) to narrow those down to a candidate set, guaranteeing the correct option falls within that set with a probability the user sets, such as 80%. If only one option remains in the set, the robot executes it directly; if several remain, it stops and asks a person. It requires no fine-tuning of the model, and was validated on mobile manipulation, tabletop rearrangement, and bimanual manipulation, keeping the number of times it has to ask for help as low as possible while still guaranteeing success.

ExampleIn a mobile-manipulation setting, the user says “put the chips in the drawer,” but there's more than one bag of chips and more than one drawer present; multiple options remain in the candidate set, so the robot stops to ask for clarification before acting.

Also called

Robots That Ask For Help, Uncertainty Alignment for LLM Planners

Related

Large Language Model · LLM-based Task Planning · Uncertainty Estimation · Hallucination · Human-in-the-Loop · SayCan

Sources

Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners (arXiv 2307.01928)
KnowNo 项目页 (Chinese)

As of

2023-11