Curriculum Learning for Embodied Planning with LLMs

Fri, 10 May 2024 00:00:00 +0000

This project explores the application of Curriculum Learning to improve the performance of GPT-2 models in Embodied Natural Language Processing tasks using the ALFWorld dataset. We developed curricula for both Action Modeling and Reinforcement Learning stages, demonstrating significant improvements in task success rates and action efficiency.

Authors: Bohan Liu, Vedaant Jain, Aarohi Gupta

Key aspects of this research include:

Developing difficulty scoring mechanisms for task demonstrations
Creating “Easy” and “Hard” curriculum sets to structure model training
Investigating the impact of curricula on model generalization across task types
Exploring the potential of few-shot learning with large language models
Demonstrating the effectiveness of a two-stage “easy-then-hard” curriculum in Reinforcement Learning

Our results show that carefully designed curricula can enhance model performance, improve generalization to unseen tasks, and increase learning efficiency in embodied AI environments.

Embodied AI | Vedaant Jain

Curriculum Learning for Embodied Planning with LLMs