Title: DAG-Plan: Constructing Directed Acyclic Dependency Graphs for Dual-Arm Cooperative Planning

Abstract:

While dual-arm robots offer the potential for significantly enhanced efficiency, they necessitate sophisticated planning strategies to handle complex tasks characterized by nonlinear sub-task dependencies. Existing approaches leveraging Large Language Models (LLMs) are constrained by a fundamental compromise: generating linear task sequences is computationally efficient but incapable of modeling parallelism or adapting to dynamic changes, whereas iterative querying methods offer adaptability but are prohibitively slow and expensive. To address this limitation, we present DAG-Plan, a novel task planning framework that utilizes a Directed Acyclic Graph (DAG) as its core representation for coordinating dual-arm operations for the first time. The central premise of this approach is that a DAG inherently represents intricate sub-task dependencies while clearly identifying opportunities for parallel execution. Within this architecture, an LLM serves a singular purpose: acting as a robust semantic parser to convert natural language instructions into a structured DAG. As the system operates, it dynamically allocates candidate nodes to the appropriate robotic arm in response to real-time environmental feedback, facilitating truly adaptive and parallel performance. Comprehensive evaluations conducted on a dual-arm kitchen benchmark demonstrate that DAG-Plan’s structured methodology significantly surpasses current paradigms. Specifically, the system achieves a 48% improvement in success rates compared to single-query linear sequence methods by effectively managing dependencies, and delivers an 84.1% boost in execution efficiency over iterative querying techniques by removing the latency associated with repeated LLM calls. Our findings indicate that adopting a principled, graph-based representation is essential for realizing efficient and reliable LLM-driven planning in complex robotic systems. Additional demonstrations and code can be accessed at https://sites.google.com/view/dag-plan.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC