LLM4RD: LLM for Reward Design

Briefly Introduction

Idea:

Originates from Eureka.

• using LLM to design reward for different kinds of sports
• in-context learning
• see details.md to know more about the project.

Environment :

The test env is mainly on SMPLOlympics

About the Code

/llm4rd：

• llm4rd.py: llm reward design loop
• /envs: change a little from SMPLOlympics/phc/env/tasks (temporarily copy boxing sports), and the obs of tasks (used for in-context learning prompting)
• /outputs: history llm output result (using LLaMA3)

smpl_envs/tasks:

• imitate the file system of SMPLOlympics. after other parts have been completed, this folder can be changed to the real SMPLOlympics file path (SMPLOlympics/phc/env/tasks)