Object navigation (ObjectNav) asks a robot to find an instance of a target object category in an unknown environment. Vision-language models (VLMs) offer the commonsense reasoning that today's robots lack, but their 3D spatial grounding is limited, and querying them as black-box oracles at every step on unstructured local observations yields redundant backtracking and brittle behavior outside simulation. This thesis argues that ObjectNav is a system-level problem rather than a single-policy learning task: semantic understanding, spatial structure, and long-horizon planning should be decoupled and handled by cooperating modules, with the VLM reasoning only at the level where it is reliable. We build such a system and carry it step by step from simulation to floor-scale, cross-embodiment deployment in the real world.

We first develop the core of this system in simulation: the robot incrementally builds a structured scene representation, the VLM reasons over it only at a high level, and geometry-based exploration handles fine-grained navigation. We then carry the system into the real world as three cooperating levels that decouple semantic reasoning, navigation planning, and motion control: the VLM provides semantically grounded guidance over the scene representation, a hierarchical room-based navigation strategy reserves that guidance for room-level decisions, and embodiment-specific motion control executes planned waypoints. Because only the lowest level depends on the robot, the same system runs on a custom-built wheeled robot, the Unitree Go2 quadruped, and the Unitree G1 humanoid. Across 190 real-world experiments it substantially improves success rate and navigates 4-5x more efficiently than existing baselines, while retaining state-of-the-art simulation performance, and, to our knowledge, it is the first system to reliably complete floor-scale, long-range ObjectNav in complex real-world environments. Real-world ObjectNav is thus solved not by a larger model or a single end-to-end policy, but by a system that balances semantic intelligence with spatial reliability and isolates embodiment-specific control from embodiment-invariant reasoning.

Master of Science in Robotics thesis. Robotics Institute Technical Report CMU-RI-TR-26-108. Thesis committee: Jean Oh (chair), Ji Zhang, Zhixuan Liu.