Eurekaverse: Environment Curriculum Generation via Large Language Models
Date read: 23rd July 2025
arXiv link
Key Points
- Uses LLMs to create a curriculum for a quadruped robotic dog to tackle obstacles.
- LLMs are used as way of tackling the problem of reliance on low dimensionality for environments.
How?
- LLM designs height grids for the agent to navigate through generating code.
- Agent-environment co-evolution: follows process of training agents and then generating code… and then back to
start.
Method:
- Generate original set of examples via a text description of the game in the prompt + some example code
- Hot sample of the LLMs outputs.
- Train a set of agents, each on a subset of these levels.
- Get best performing agent on all levels… the levels used to train this agent must be strong.
- Give each of these tasks to the LLM with the task description and ask for it to make the task harder.
- Repeat process but with set of the best performing agent