An Embodied Collaborative Search Strategy for Multi-Scale Visual Feature Targets
Yucheng Chen, Yanze Zhu, Chunxing Guan, Jilong Zheng · 2024
Compared to canines, quadruped robots offer significant advantages, including lower training costs, ease of large-scale deployment, and reduced maintenance burdens. As quadruped robot technology has gradually matured, most studies have focused on the kinematics and navigation of quadruped robots with limited research on collaborative search tasks involving canines and biomimetic robots. To develop a quadruped robot guidance system that enhances search success rates, this study addresses the demand for collaborative operations between canines and quadruped robots. This study proposes a multi-scale visual target perception mechanism, which is based on confidence region estimation. This mechanism strengthens adaptability to target scale variations in complex scenes, enabling more accurate target recognition. Inspired by traditional tracking dog training methods, we propose an embodied collaborative and efficient search strategy. We evaluate the strategy by testing the success rates of target search tasks in both daytime and nighttime indoor environments. The embodied collaborative search strategy improves, recognition speed by approximately 62.5% in dark environments and 35.3% in illuminated environments, compared to traditional single canine search tasks and the overall task completion rate is also improved. This was shown in a number of different experimental scenarios that took place in different spatiotemporal settings.