# Inner Monologue: Embodied Reasoning through Planning with Language Models

**Authors:** Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, Brian Ichter  
**arXiv:** [2207.05608](https://arxiv.org/abs/2207.05608)  
**Date:** 2022-07-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.05608v1/

## Key points

- Large Language Models (LLMs) have potential in domains beyond natural language processing
- LLMs used in embodied environments face additional challenges
- The study investigates the extent to which LLMs can reason over different sources of feedback provided through natural language without additional training
- Feedback from the environment helps LLMs develop an inner monologue that enhances their ability to process information and plan actions in robotic control scenarios
- Various forms of feedback, including success detection, scene description, and human interaction, are explored
- Incorporating closed-loop language feedback significantly improves high-level instruction completion across three domains: simulated and real table-top rearrangement tasks and long-horizon mobile manipulation tasks in a kitchen environment
- Improvement is observed both in simulations and real-world experiments
- Natural language feedback enhances reasoning capabilities in embodied contexts
- Enabling LLMs to form an inner monologue through environment feedback improves task completion based on high-level instructions across different domains

## Summary

Recent works have demonstrated the potential of Large Language Models (LLMs) in domains beyond natural language processing, such as planning and interaction for robots. These tasks require the agent to understand various semantic aspects of the world, including available skills, their impact on the environment, and how changes in the world are represented in language. However, LLMs used in embodied environments face additional challenges. They not only need to determine what skills to employ but also how and when to execute them, with answers that may change over time based on the agent's choices. In this study, the authors investigate the extent to which LLMs can reason over different sources of feedback provided through natural language without any additional training. They propose that by leveraging feedback from the environment, LLMs can develop an inner monologue that enhances their ability to process information and plan actions in robotic control scenarios. The researchers explore various forms of feedback, including success detection, scene description, and human interaction. The findings reveal that incorporating closed-loop language feedback significantly improves high-level instruction completion across three domains: simulated and real table-top rearrangement tasks and long-horizon mobile manipulation tasks in a kitchen environment. This improvement is observed both in simulations and real-world experiments. Overall, this research highlights the potential of using natural language feedback to enhance reasoning capabilities in embodied contexts. By enabling LLMs to form an inner monologue through environment feedback, they become more proficient at processing information and planning actions in robotic control scenarios. These findings contribute to advancing the field of robotics by improving task completion based on high-level instructions across different domains.