| （英） |
Cost-effective path planning is essential for unmanned systems, particularly considering the challenges of recharging during prolonged outdoor operations. Despite its significance, this area has been largely overlooked in prior research. Traditional path planning methods are proposed aiming at obstacle avoidance, often resulting in cost-inefficient paths. Meanwhile, large multimodal models (LMMs) exhibit exceptional contextual understanding but remain underutilized in path planning due to their limited spatial awareness.
To bridge this gap, we present Llava-Planner, a large multimodal model designed for cost-effective path planning. Llava-Planner explores a novel approach to leverage LMMs for path planning. This paper also proposes three carefully designed pretraining tasks to further enhance the spatial awareness of Llava-Planner, including: terrain description generation, identification of the coordinates of the start and end point coordinates, and obstacle avoidance prediction.
Experimental evaluations on a newly created grid map dataset demonstrate that LMMs can effectively perform path planning with well-crafted prompts, demonstrating their potential as a robust and efficient solution for robot path planning and navigation. However, there are some random errors found in Llava-Planner generated path points, highlighting the need for human-in-the-loop integration as a potential future direction to enhance reliability and performance. |