UML-CoT: Structured Reasoning and Planning with Unified Modeling Language for Robotic Room Cleaning

Hongyu Chen, Guangrun WangarXiv 2025 · DOI · arXiv

Introduction

Large Language Models (LLMs) have significantly advanced robotic reasoning through Chain-of-Thought (CoT) prompting. However, text-based reasoning often lacks the precision required for physical interaction, leading to logical inconsistencies and "shallow" world models. Robotic planning requires a representation that is more formal than free-form text but more expressive than simple scene graphs. To address this, the authors propose UML-CoT, a framework that leverages the standardized semantics of software engineering to guide robotic "thought."

A cluttered bedroom environment serving as the input for the cleaning task.

Methods

UML-CoT replaces unstructured text with standardized Unified Modeling Language (UML) diagrams. By using formal diagrams, the model can capture complex relationships like object inheritance and procedural control flows that are often lost in natural language.

Structured Representation

The framework utilizes two primary UML components generated as PlantUML code:

  1. UML Class Diagrams: Used for symbolic reasoning to capture object attributes, hierarchical relationships (e.g., a "dirty plate" as a type of "dish"), and spatial context.
  2. UML Activity Diagrams: Used to generate executable cleaning plans, encoding sequential, conditional, and parallel control flows.
The UML-CoT approach vs. unstructured text CoT, showing the transition from image input to structured class and activity diagrams.

Training Pipeline

The system, built on the InternVL 2.5 backbone, follows a three-stage training process:

  • Supervised Fine-Tuning (SFT): The model is trained on high-quality, human-verified UML reasoning traces.
  • Reinforcement Learning (RLFT): Using Group Relative Policy Optimization (GRPO), the model is rewarded based on the success of the final plan, encouraging coherent intermediate reasoning.
  • Answer-Only GRPO: Further optimization on larger datasets ensures generalization beyond the initial training samples.

The MRoom-30k Benchmark

To evaluate the framework, the researchers introduced MRoom-30k, a large-scale dataset featuring over 30,000 high-quality images of cluttered real-world environments (kitchens, bedrooms, living rooms). This benchmark provides a diverse set of "messy" scenarios to test the robustness of robotic cleaning plans.

Key Results and Insights

UML-CoT demonstrates superior performance over traditional text-based and graph-based methods:

  • Success Rate: Achieved a 67.44% success rate, significantly outperforming text-based CoT (62.80%) and Graph of Thoughts (55.79%).
  • Interpretability: The use of valid UML allows the robot's reasoning process to be visually rendered, making it easier for humans to verify and debug.
  • Latent Reward Propagation: Rewarding the final plan effectively improves the quality of the intermediate "thoughts," proving that structured representations facilitate better logical backpropagation.

Conclusion

UML-CoT offers a promising path toward more reliable and interpretable robotic reasoning. By formalizing the "Chain-of-Thought" into structured diagrams, the framework significantly improves a robot's ability to navigate and clean complex, real-world environments. Future work will focus on improving cross-task generalization and deploying these structured plans on physical hardware.

Summary updated

BibTeX
@misc{chen2025uml,
  author = {Hongyu Chen and Guangrun Wang},
  title = {{UML-CoT: Structured Reasoning and Planning with Unified Modeling Language for Robotic Room Cleaning}},
  year = {2025},
  eprint = {2509.22628},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.22628},
}
Figure