RoboPARA: Dual-Arm Robot Planning with Parallel Allocation and Recomposition Across Tasks

Dual-arm robots play a crucial role in improving efficiency and flexibility in complex multitasking scenarios.While existing methods have achieved promising results in task planning, they often fail to fully optimize task parallelism, limiting the potential of dual-arm collaboration.To address this issue, we propose RoboPARA, a novel large language model (LLM)-driven framework for dual-arm task parallelism planning.RoboPARA employs a two-stage process: (1) Dependency Graph-based Planning Candidates Generation, which constructs directed acyclic graphs (DAGs) to model task dependencies and eliminate redundancy, and (2) Graph Re-Traversal-based Dual-Arm Parallel Planning, which optimizes DAG traversal to maximize parallelism while maintaining task coherence.In addition, we introduce the Cross-Scenario Dual-Arm Parallel Task dataset (X-DAPT dataset), the first dataset specifically designed to evaluate dual-arm task parallelism across diverse scenarios and difficulty levels.Extensive experiments demonstrate that RoboPARA significantly outperforms existing planning methods, achieving higher efficiency and reliability, particularly in complex task combinations.Our code is publicly available at https://github.com/AiDuanshiying/RoboPARA . To tackle the issue, we propose RoboPARA, a novel LLM-driven dual-arm robot planning framework. RoboPARA ensures both task correctness and optimal arm utilization through a two-stage architecture: Dependency Graph-based Planning Candidates Generation and Graph Re-Traversal-based Dual-Arm Parallel Planning. In the former stage, RoboPARA processes user instructions by utilizing a local memory module, supported by retrieval augmented generation (RAG) (Gao et al., 2023), to obtain detailed procedural knowledge relevant to the task. This retrieved information is integrated into a structured prompt, enabling the LLM to generate a directed acyclic graph (DAG) with correction that outlines the step dependencies. In the later stage, the generated DAG is analyzed and optimized by scheduling algorithms to fully exploit the collaborative potential of dual-arm robots. Together, these two stages form a tightly integrated pipeline that enables RoboPARA to generate task plans that not only ensure accuracy and feasibility but also optimize dual-arm collaboration. To validate the effectiveness of RoboPARA, we further propose the Cross-Scenario Dual-Arm Parallel Task dataset (X-DAPT dataset), the first dataset designed to evaluate dual-arm task planning parallelism. Our dataset includes more than 1,000 task packages across 10 key scenarios, each divided into three difficulty levels. Experiments demonstrate that RoboPARA achieves outstanding performance, especially in the most complex task combinations. Compared to existing methods, RoboPARA exhibits over 4.5× parallel and collaborative steps in average, resulting in a 30% to 50% reduction in execution time, significantly improving efficiency. Moreover, RoboPARA achieves 34% higher success rate than the average of other methods in challenging tasks, ensuring greater reliability and stability in task execution. Our contributions are summarized as follows: • New Task: We propose a new dual-arm task planning problem, the Dual-Arm Cooperative Scheduling Problem, specifically designed to optimize parallelism and improve multitasking efficiency in real-world scenarios. Our work establishes a new task and objective, pushing the boundaries of dual-arm robotic manipulation. • New Dataset: We design the X-DAPT dataset, the first dataset tailored for evaluating dual-arm task parallelism. It includes over 1,000 task packages across 10 diverse scenarios with difficulty levels, offering a comprehensive benchmark. • New Method: We propose RoboPARA, an LLM-driven dual-arm planning framework using a two-stage process: Dependency Graph-based Planning and Graph Re-Traversal. It ensures efficient, reliable, and highly parallel task execution, achieving state-of-the-art task success rates and execution efficiency on X-DAPT dataset and real-world scenarios. RELATED WORK Dual-arm Robot Manipulation. Dual-arm manipulation plays a critical role in enabling robots to perform complex, high-precision tasks in various domains (Zhao et al., 2025; Yoshida et al., 2023) . Current approaches to dual-arm manipulation can be broadly categorized into two main paradigms: End-to-end methods train a self-contained model to directly map from task descriptions to robotic trajectory, such as HybridVLA (Liu et al., 2025b), DA-VIL (Karim et al., 2024 ), R3M (Nair et al., 2022), Long-VLA (Fan et al., 2025). However, such methods have limited scalability and flexibility for long-horizon tasks, and generalizing to new tasks or environments often requires extensive retraining. Compositional skill learning involves developing multiple meta-skills that can be combined to accomplish complex tasks, offering greater flexibility and reusability (