AGIBOT's HarnessVLN posts 76% on HM3D-v2 training-free
TL;DR
- HarnessVLN reports success rates of 60.8% on R2R, 53.9% on RxR, 76.0% on HM3D-v2 and 59.3% on HM3D-OVON, all zero-shot and training-free.
- Its abstract claims gains of 5.8, 12.1, 1.6 and 9.1 percentage points over prior training-free state-of-the-art on those four benchmarks.
- Authors from Nanjing University, AGIBOT and Tsinghua deploy the same Agent Harness on a humanoid robot for instruction-following and object-goal navigation.
HarnessVLN, a zero-shot navigation framework from researchers at Nanjing University, AGIBOT and Tsinghua University, reports success rates of 60.8% on R2R, 53.9% on RxR, 76.0% on HM3D-v2 and 59.3% on HM3D-OVON without task-specific training. The paper on Hugging Face says those numbers improve on prior training-free state-of-the-art results by 5.8, 12.1, 1.6 and 9.1 percentage points respectively.
The central move is an Agent Harness sitting between a multimodal LLM planner and the environment. The abstract describes it as a component that "coordinates perception, retrieval, grounding, navigation, recovery, and termination through a unified tool interface," backed by a hierarchical event memory and a persistent Spatiotemporal Graph that feed validated feedback into subsequent planning steps.
A replaceable Navigation Executor turns validated targets into motion, and the authors argue the same Harness protocol supports both instruction-following and object-goal navigation. The paper adds that "humanoid deployment further demonstrates its applicability to both tasks in real-world environments," though the abstract does not name the humanoid platform, the underlying MLLM, or per-episode inference cost.
It lands alongside two other harness-shaped agent papers we logged in the agents feed the same day.
Originally reported by huggingface.co
Read the original article →Original headline: HarnessVLN Paper Unifies Training-Free Embodied Navigation, Hits 76% on HM3D-v2