[paper-review] Text2Reaction : Enabling Reactive Task Planning Using Large Language Models

RA-L, 2024. [Paper]

Zejun Yang , Li Ning, Haitao Wang, Tianyu Jiang, Shaolin Zhang, Shaowei Cui, Hao Jiang, Chunpeng Li, Shuo Wang and Zhaoqi Wang

May. 05.

Fig. 1: Overview of Text2React.

Title:

Text2Reaction : Enabling Reactive Task Planning Using Large Language Models (R-AL, 2024)

Summary:

They propose Text2Reaction, an LLM-based framework enabling robots to continuously reason and update plans according to the latest environment changes.

Contribution:

Fig. 2: Flowchart of Text2React.

Fig. 3: Reasoning Steps of Text2React.

• They present the Re-planning Prompt, which informs LLMs the basic principles of re-planning.
  • It fosters the gradual development of a current plan to a new one in a three-hop reasoning manner: cause analysis, consequence inference, and plan adjustment
• OffPlanner: an LLM-based planner that generates initial plans
• On-Planner: another planner, which updates plans under the guidance of the re-planning prompts

Thoughts:

• Re-planning is an important part of the reactive robot.
  • They showed an LLM-based framework capable of comprehensively analyzing various feedback and continuously re-planning in response to environment changes.
• They propose new evaluation metrics for the success rate of task replanning: Executability Rate(ER), Success weighted by Path Length(SPL).