HarnessVLN: Unifying Training-Free Embodied Navigation through an Agent Harness

HarnessVLN. Unifying Training-Free Embodied Navigation through an Agent Harness. By Yang Chen and nine others, from Nanjing University, AgiBot, and one other institution. Published on the fourteenth of September. No code repository is linked to it.

Embodied navigation asks an agent to interpret what it sees, accumulate spatial knowledge, and act, whether it is following instructions or searching for an object. Training-based methods face challenges in generalization. Training-free methods lean on multimodal large language models, or MLLMs, instead. But they often lack any way to reconcile a proposed action with three things that should constrain it: the spatial evidence, progress on the task, and failures in execution. A model can propose a move that the map contradicts, and nothing catches it.

HarnessVLN is a zero-shot, training-free framework built around an Agent Harness. The harness coordinates perception, retrieval, grounding, navigation, recovery and termination through one unified tool interface. Crucially, it validates the planner's proposals against spatial evidence, geometric feasibility and subgoal consistency, and it feeds structured tool feedback back into the next decision.

Two kinds of memory support that checking. Hierarchical event memory tracks task progress and execution history, so the agent knows where it is in the task. A persistent Spatiotemporal Graph holds reusable spatial evidence along with annotations of past failures, which are then used for verification and for recovery. Finally, a replaceable Navigation Executor turns validated targets into executable motions. Because the executor can be swapped, the same harness protocol serves both instruction-following and object-goal navigation.

Success rates are sixty point eight percent on R2R, fifty-three point nine percent on RxR, seventy-six percent on HM3D-v2, and fifty-nine point three percent on HM3D-OVON. Each surpasses the prior training-free state of the art. Four benchmarks, one harness protocol, and no training on any of them, which is the zero-shot claim the framework was built around. A humanoid deployment further demonstrates both tasks working in real-world environments.

The design separates proposing from checking, which is the gap the problem statement pointed to. The language model proposes where to go, and the harness decides whether that proposal agrees with what the agent has actually seen and done. Keeping the executor replaceable is what lets one protocol cover two different navigation tasks.