Scale-Plan: Scalable Language-Enabled Task Planning for Heterogeneous Multi-Robot Teams

A team of researchers including Piyush Gupta, Sangjae Bae, Jiachen Li, and David Isele has introduced Scale-Plan, a novel framework designed to solve a core challenge in robotics: scalable task planning for teams of different robots. Traditional symbolic planners require labor-intensive, manual problem specifications, while recent Large Language Model (LLM)-based approaches often produce unreliable "hallucinated" plans that aren't properly grounded in the real environment. Scale-Plan bridges this gap by using LLMs not to generate the final plan directly, but to intelligently guide a structured, reliable planning process.

Given a high-level instruction and a formal Planning Domain Definition Language (PDDL) specification, Scale-Plan first constructs an action graph capturing the domain's structure. It then employs "shallow" LLM reasoning to steer a graph search, identifying only the minimal subset of relevant actions and environmental objects needed for the task. By filtering out the vast majority of irrelevant perceptual data *before* the planning step, the system dramatically reduces complexity. This allows for efficient decomposition of tasks, allocation to specific robots, and generation of long-horizon plans. The team also released MAT2-THOR, a cleaned benchmark based on AI2-THOR, to provide a reliable testbed for future multi-robot planning systems. In evaluations on complex tasks, Scale-Plan consistently outperformed both pure-LLM and hybrid LLM-PDDL baseline methods across all metrics, demonstrating superior scalability and reliability.