Hypothesis-driven Model Expansion under Uncertainty
for Open-World Robot Planning

Robotics: Science and Systems (RSS) 2026

Anxing Xiao, Hanbo Zhang, Tianrun Hu, David Hsu

School of Computing & Smart Systems Institute

National University of Singapore

Paper (Coming Soon) Code (Coming Soon) Video

Key Messages: In open-world settings, robots should continuously expand their models of the environment on the fly. The robot should also plan ahead under the uncertainty introduced by such model expansion.

Framework Overview

Framework overview of HUME
The robot iteratively generates hypotheses to expand its model, plans with a model augmented by uncertain hypotheses, and executes actions to verify and update these hypotheses using newly acquired observations.

HUME tackles open-world robot planning by treating missing knowledge as uncertain, testable hypotheses rather than fixed facts. Given a task instruction and an incomplete world model, the system uses foundation models to propose task-relevant hypotheses about object locations, attributes, and action effects, then plans actions that both make progress toward the goal and verify those assumptions through interaction. This loop of hypothesis generation, planning, execution, and update allows the robot to continually expand its world model while remaining grounded in real observations.

Real-World Environment

Real-world household setup with Fetch robot
The real-world setup includes a kitchen, living room, and dining area, with a Fetch mobile manipulator operating under substantial task openness: The robot only knows the locations but doesn't know the objects.

Tasks

Deliver a zero-sugar drink to the table

Place the remote with a red button into the cabinet

Move the smiley-face mug to the fridge

Serve a heated chicken burger on the coffee table

Throw away the blue-floral bowl and turn off the livingroom light

Defrost food for 40 seconds (microwave operation)