Researchers have published HarnessVLN, a zero-shot, training-free framework for embodied navigation, on Hugging Face. The work targets a persistent tension in the field: training-based navigation agents struggle to generalize, while training-free approaches built on multimodal large language models (MLLMs) often lack a mechanism to reconcile proposed actions with spatial evidence, task progress, and execution failures.
The proposed solution is an Agent Harness that coordinates perception, retrieval, grounding, navigation, recovery, and termination through a unified tool interface. According to the paper, the Harness validates planner proposals against spatial evidence, geometric feasibility, and subgoal consistency, and feeds structured tool feedback back into subsequent decisions.
Two memory components support this loop. A hierarchical event memory tracks task progress and execution history, while a persistent Spatiotemporal Graph stores reusable spatial evidence and failure annotations for verification and recovery. A replaceable Navigation Executor then converts validated targets into executable motions, letting the same Harness protocol serve both instruction-following and object-goal navigation.
The authors report success rates of 60.8% on R2R, 53.9% on RxR, 76.0% on HM3D-v2, and 59.3% on HM3D-OVON, which they say surpass prior training-free state-of-the-art results. A humanoid deployment is also presented as evidence that the approach applies to both task types in real-world environments.
The project page is available at harnessvln.netlify.app. The paper appears alongside related recent work on agentic memory and neuro-symbolic navigation surfaced by Hugging Face's recommendation bot.