LLM-iTeach for Robot Manipulation with RLBench

Figure: Comparison of General IIL framework and our proposed LLM-iTeach. In IIL, a human teacher observes the agent’s actions and provides timely feedback. In contrast, LLM-iTeach first prompts the LLM to encapsulate its reasoning hierarchically into CodePolicy, and then provides feedback to the agent in an evaluative or corrective manner through the similarity-checking mechanism designed in LLM-iTeach.

Description

This repository contains the code used for the experiments presented in my master's thesis, where we evaluate LLM-iTeach and Behavior Cloning (BC) on robot manipulation tasks from the RLBench benchmark. The experiments follow the CEILing framework to ensure comparability.

⚠️ Note: The code was used as-is with additional computer specific configurations to generate the results in the paper. Compatibility on individual systems is not guaranteed, and adjustments may be required depending on your environment. Please feel free to reach out if you need assistance.

Key Notes

• LLM Configuration: The code requires access to an LLM and must be configured appropriately in LMP.py.
• Prompt Design: Prompt fragments live in prompts/ and are assembled hierarchically in lmp.py.
• Generated Policies: The codepolicies/ directory contains LLM-generated control code used by the agent in task execution.
• Camera Support: The current implementation does not fully support the dual-camera setup in RLBench.

📄 Thesis Paper

🔗 Link to the paper