Dual-arm coordination is a fundamental problem in humanoid nursing robot. Large language model (LLM)-driven dual-arm collaboration is gradually becoming a research hotspot in this field. However, the single-thread LLM task planner lacks the ability of co-scheduling, which leads to poor efficiency in nursing robot. To cope with the problem, this letter proposed a multi-agent LLM solution for the task planning of nursing robot, named DABICO. The framework constructs dual agent systems (left-arm and right-arm) at the levels of communication and decision-making, as well as ensuring a single robot entity. Moreover, we construct corresponding communication mechanism and dialogue protocol to promote the information exchange between the two agents. Finally, validation feedback system is proposed to ensure that the sub-task of each robot arm can be executed successfully. A large set of experiments show that, compared to the single-thread LLM task planner, the DABICO framework is more advantageous when dealing with the bimanual coordination tasks. DABICO is able of accomplishes reasoning rapidly, reducing Replan metrics by $\mathbf{90\%}$ on average, and the improvement with respect to Success rate is $\mathbf{11\%}$ on average. Finally we demonstrate DABICO in real-world medicine organization experiment on a dual-arm nursing robot.

Nursing robots are designed to serve users, and their ability to interact with humans, as well as to make task-related decisions and decompositions based on such interactions, is a fundamental prerequisite for autonomous execution of nursing tasks. Large language models offer an effective approach to facilitating human–robot interaction. However, their global perspective can lead to confusion or reduced precision when coordinating the execution of tasks by a dual-arm robot, often generating execution sequences that are inconsistent with real-world conditions. To address this challenge, this study proposes a multi-agent framework, wherein each arm of the nursing robot is conceptualized as an independent agent. Through the application of geometric constraints, these agents maintain appropriate relative positional relationships and achieve coordinated collaboration via a large language model. This approach enhances the task planning capabilities of the robot and improves its efficiency in delivering nursing services.

Large language models (LLMs)-driven robotic systems have become a focal point. However, current methodologies predominantly rely on auxiliary components such as pre-defined motion primitives or pre-trained skills to facilitate physical environment actions, which constitutes the primary limitation for such approaches. This study presents the SeIn framework that breaks through this limitation by leveraging prompt-conditioned sequential generation capability of LLMs to directly output control actions for dual-arm nursing robot—specifically, sequences of 14-DoF target joint positions. Firstly, the LLMs were grounded in a MuJoCo physics environment, enabling the dual-arm nursing robot to generate control actions while receiving corresponding environmental observations. Subsequently, task-specific textual prompts were designed to guide the LLMs in task reasoning, observation processing, and generation of target joint position sequences. Finally, the framework was evaluated through comprehensive MuJoCo simulations, including: ablation experiments on prompt design; expressive action experiments across seven action types spanning three operational categories; performance evaluation experiments querying four distinct LLMs at 4- and 60-s execution durations; and a functional action experiment involving Reach-to-Object. Experimental results demonstrated that SeIn framework achieved 100% success rate in generating joint sequences for actions using task-specific text prompts and pre-trained LLMs (specifically, GPT-4o and DeepSeek-V3.1), without requiring auxiliary components.

Effective bimanual collaboration represents a significant research objective for Large Language Model (LLM)-driven nursing robots. However, current approaches are constrained by critical limitations: the single-thread LLM task planner (ST-Planner) lacks co-scheduling, while the conventional multi-agent framework (DABICO) suffers from inefficient information interaction between agents, consequently compromising performance in bimanual collaboration tasks. To overcome these limitations, this study introduces Role-playing Dual Agents (RoDA), a novel dual-agent collaboration framework augmented by LLM Role-Playing. This framework implements the nursing robot as an LLM-based dual-agent system wherein each agent assumes the role of either the left or right arm. Through meticulously crafted contextual prompts explicitly defining specific identity attributes and conversational protocols for each limb, these agents facilitate high-quality collaborative dialogue reflective of their designated roles. Evaluation of RoDA was conducted through four MuJoCo simulation scenarios encompassing all four categories of bimanual collaboration, diverse task sequences, and different degrees of workspace overlap. The experimental results demonstrate that the role-playing mechanism enhances dialogue normalization, accuracy, and information richness. This high-quality interaction enables superior task planning performance, with RoDA surpassing both the ST-Planner and the baseline DABICO framework. Specifically, RoDA achieved a \((100%)\) success rate across all scenarios, exhibiting an average performance improvement of \((16.3%)\) in the Step metric and \((70.2%)\) in the Restep metric relative to DABICO. Furthermore, enhancement of performance is achieved without the requirement for LLM fine-tuning, offering advantages including flexibility, immediacy, and low development cost. Finally, RoDA was demonstrated through practical experiments on a dual-arm nursing robot.