Formal and Natural Language assisted Curriculum Generation for Reinforcement Learning Agents

Reinforcement Learning (RL) has proven successful in learning behaviors for artificial agents and robots when the transition dynamics of the environment are unknown. Despite this progress, many sequential decision making tasks are prohibitively expensive to learn. For my research, I intend to utilize and synthesize existing symbolic knowledge available to supplement the RL techniques for improved efficiency and faster learning progress. This symbolic information be in the form of formal language specifications (such as LTL) or in the form of natural language derived using Large Language Models (LLMs). I have developed various methods and frameworks that propose novel techniques in the curriculum learning domain to improve the learning efficiency of RL agents. I further want to implement these techniques on physical manipulator robot and show its efficacy for solving problems in the real world.