DRP: A Decomposition-Reflection-Prediction Framework for Long-Horizon Robot Task Planning using Large Language Models

Large language models have demonstrated powerful reasoning capabilities, and their integration with robotics has revolutionized human-computer interaction and automated task planning. However, LLMs are unaware of environmental knowledge and possible state changes in the environment during planning, which makes the generated tasks unexecutable, particularly when dealing with complex long-horizon tasks involving crowded objects and dynamic relations. In this paper, we propose a LLM-based robot task planning framework with support for environmental knowledge injection, which is called DRP(Decomposition-Reflection-Prediction). The DRP framework combines LLMs with rule-based task decomposition, multi-perspective reflection and environmental prediction to generate admissible actions for complex long-horizon tasks. We only leverage few-shot prompting to implement our framework, which avoids the need for additional model training work. Experiments on VirtualHome household task dataset show that the task plans generated by our method have improved the executability by 25. 23%, the subgoal success rate by 64. 29%, and the success rate by 58. 06%, in comparison to state-of-the-art baseline methods. The complete code of our framework has been made public at https://github.com/lab-bj/taskplanning

Authors

• Zhaowen Zheng
• Zhuofeng Zhao
• Haocen Wang
• Jing Wang

Keywords

• Training
• Human computer interaction
• Codes
• Large language models
• Reflection
• Cognition
• Planning
• Intelligent robots
• Language Model
• Task Planning
• Baseline Methods
• Environmental Predictors
• Plan Generation
• Environmental Conditions
• Training Set
• Action Plan
• Sequence Of Actions
• Task Completion
• Introspection
• Corrective Actions
• Activation Step
• Rating Task
• Final Goal
• Objective Conditions
• Object Relations
• Class Instances
• Prediction Rule
• Coffee Table
• Cupcake
• Scene Graph
• Linearizable
• Impact Of Sequence
• Task Setting
• Task Execution