Researchers introduce HarnessVLN, a zero-shot, training-free agent harness that unifies instruction-following and object-goal embodied navigation by coordinating perception, memory, and execution tools through a shared interface that validates planner proposals against spatial evidence, geometric feasibility, and subgoal consistency before dispatch. The harness maintains a hierarchical event memory and a persistent spatiotemporal graph to track task progress and contextualize failures, using structured execution feedback to guide subsequent planning and recovery. It achieves success rates of 60.8%, 53.9%, 76.0%, and 59.3% on the R2R, RxR, HM3D-v2, and HM3D-OVON benchmarks respectively, outperforming prior training-free methods, and is also demonstrated on a real-world humanoid robot.
