Hugging Face Daily PapersYang Chen, Lirong Che, Zhenyu Huang1 min readpaperadvanced
HarnessVLN: Unifying Training-Free Embodied Navigation through an Agent Harness
Summary
HarnessVLN introduces a zero‑shot, training‑free embodied navigation framework that wraps a multimodal LLM in an "Agent Harness" – a tool‑based protocol that validates planner actions against spatial evidence, tracks progress with hierarchical event memory, and maintains a persistent spatiotemporal graph for recovery. The system works for instruction‑following and object‑goal tasks, achieving 60.…
- Agent Harness provides a unified interface (perception, retrieval, grounding, navigation, recovery, termination) that lets a frozen MLLM propose actions which are then programmatically verified for geometric feasibility…
- Hierarchical event memory records task progress and execution history, while a persistent spatiotemporal graph stores reusable spatial evidence and failure annotations for later verification and recovery.
- A plug‑in Navigation Executor translates validated targets into low‑level motions, enabling the same harness to support both instruction‑following (R2R, RxR) and object‑goal (HM3D‑v2, HM3D‑OVON) navigation.
- Zero‑shot performance surpasses prior training‑free baselines: 60.8 % (R2R), 53.9 % (RxR), 76.0 % (HM3D‑v2), 59.3 % (HM3D‑OVON).
Training‑free navigation sidesteps costly data collection and domain‑specific fine‑tuning, but prior methods struggled with grounding LLM proposals in concrete spatial constraints. HarnessVLN’s tool‑driven verification loop and persistent spatial memory give a practical recipe for reliable, zero‑sh…
8/10
