GroundedPlanBench: 空间基础的长时间任务规划用于机器人操作

The study introduces a new benchmark framework, GroundedPlanBench, aimed at addressing long-horizon task planning challenges in robot manipulation. Existing vision-language models (VLMs) typically separate the decision-making process into two steps: generating a natural language plan and translating it into executable actions. This approach often leads to failures in decision-making during practical applications. GroundedPlanBench enhances VLMs' task planning capabilities by incorporating spatial information, enabling more effective handling of long-horizon tasks in complex environments. This improvement boosts the efficiency and accuracy of robots operating in dynamic scenarios, providing a significant reference standard for future research.