Do It for HER: First-Order Temporal Logic Reward Specification in Reinforcement Learning (Extended Version)

Researchers developed a framework that allows complex tasks for AI agents to be specified using first-order logic, a more expressive form of instruction. This eliminates the need for manual, low-level coding of rewards in reinforcement learning. The method combines logical specification with a tailored Hindsight Experience Replay technique to overcome sparse rewards. Experiments in continuous-control settings show it enables the natural specification of intricate goals that were previously difficult to encode.