DART-LLM: Dependency-Aware Multi-Robot Task Decomposition and Execution for Construction Robotics Using Large Language Models

Abstract

This study proposes DART-LLM: Dependency-Aware Multi-Robot Task Decomposition and Execution for Construction Robotics using Large Language Models. Construction sites require robust multi-robot coordination that is reliable, human-verifiable, and deployable on edge devices with unstable network connectivity. However, existing multi-robot coordination approaches lack human-verifiable task decomposition methods suitable for high-stakes construction environments, where task failures can result in severe safety incidents and existing cloud-dependent solutions are impractical for edge deployment in remote sites with unreliable connectivity. DART-LLM addresses these limitations by employing Directed Acyclic Graphs (DAGs) to model task dependencies, enabling decomposition of natural language instructions into well-coordinated subtasks with transparent, human-interpretable visualization.The system provides intuitive DAG visualization for human verification and supports iterative task refinement through interactive dialogue.Experimental evaluation on natural language instructions across three complexity levels demonstrates that DART-LLM achieves higher success rates compared to existing baselines.Ablation studies confirm that explicit dependency modeling significantly improves small-model performance. The fine-tuned Llama-1B runs on resource-constrained Jetson devices, and multi-robot coordination is validated on physical robots. All module code and the fine-tuned model are publicly available.

Experimental Results

Note: The simulation results shown below represent the first 6 examples from the YongdongWang/dart_llm_tasks dataset. We have also conducted real robot experiments with 2 examples. The natural language prompts displayed below each demonstration video represent the original task instructions provided to the DART-LLM system. For real robot experiments, annotations such as "(Real Robot) - * view" are added for clarity and are not part of the input prompts.

"L1-T001: Dump truck 1 goes to the puddle for inspection, after which all robots avoid the puddle."

"L1-T002: Send the Excavator 1 to the obstacle, and perform excavation."

"L2-T001: Send Excavator 1 and Dump Truck 1 to the soil area; Excavator 1 will excavate and unload, followed by Dump Truck 1 proceeding to the puddle for unloading."

"L2-T002: Move Excavator 1 and Dump Truck 1 to soil area 2; Excavator 1 will excavate and unload, then Dump Truck 1 returns to the starting position to unload."

"L3-T001: Excavator 1 is guided to the obstacle to excavate and unload to clear the obstacle, then excavator 1 and dump truck 1 are moved to the soil area, and the excavator excavates and unloads. Finally, dump truck 1 unloads the soil into the puddle."

"L3-T002: Excavator 1 goes to the obstacle to excavate and unload to clear the obstacle. Once the obstacle is cleared, mobilize all available robots to proceed to the puddle area for inspection, then all robots avoid the puddle. Then Excavator 1 performs excavation and unloading while Dump Truck 2 returns to the starting position."

Real Robot Experiments

"L1-T001: Dump truck 1 goes to the puddle for inspection, after which all robots avoid the puddle. (Real Robot) - Top view"

"L1-T001: Dump truck 1 goes to the puddle for inspection, after which all robots avoid the puddle. (Real Robot) - Camera view"

"L2-T001: Send Excavator 1 and Dump Truck 1 to the soil area; Excavator 1 will excavate and unload, followed by Dump Truck 1 proceeding to the puddle for unloading. (Real Robot) - Top view"

"L2-T001: Send Excavator 1 and Dump Truck 1 to the soil area; Excavator 1 will excavate and unload, followed by Dump Truck 1 proceeding to the puddle for unloading. (Real Robot) - Camera view"