HINT-VLM: Humanoid Interactive Navigation over Rough Terrain with Embodiment-Grounded Vision-Language Model
HINT-VLM: Humanoid Interactive Navigation over Rough Terrain with Embodiment-Grounded Vision-Language Model
This research introduces HINT-VLM, a hybrid planning framework for humanoid robots navigating cluttered outdoor rough terrain. The method combines semantic priors from a Vision-Language Model (VLM) with embodied robot interaction data via a Gaussian process (GP) residual model to correct VLM-derived movability estimates. These embodiment-grounded probabilistic movability estimates are integrated into hierarchical uncertainty-aware planning, enabling humanoids to jointly leverage locomotion and manipulation for efficient and safe navigation. This represents the first solution to Navigation Among Movable Obstacles (NAMO) in rough terrain through predictive movability reasoning and probabilistic VLM integration.
The paper proposes a hybrid planning framework combining Vision-Language Model semantic priors with Gaussian process residual models to estimate object movability for humanoid navigation in cluttered outdoor rough terrain, addressing the limitation of prior NAMO approaches that focus on flat environments or reactive interaction.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org