Anticipatory Reinforcement Learning for Improving Continual Planning in Persistent Environments

Abstract

Robots deployed in persistent environments, such as homes and restaurants, are increasingly asked to carry out long sequences of tasks where the world does not reset between instructions. A myopic agent that optimizes only for the immediate task can inadvertently leave the environment in a state that raises the expected cost of future tasks. Anticipatory planning mitigates this by preparing for predictable future routines. However, existing anticipatory planning methods rely on expensive, planner-generated offline datasets and are limited to a single-task lookahead objective, so they fail when preparatory actions must be amortized over a longer horizon. In this thesis, we propose a neurosymbolic approach that learns an anticipatory value function online via Reinforcement Learning. This creates a multi-task credit horizon without requiring a planner-generated anticipatory cost dataset. We deploy this value as a terminal heuristic to guide a cost-bounded symbolic search, and evaluate it in persistent two-room and three-room restaurants. Our value-guided anticipatory agent reduces total sequence cost by 31.8% (two-room) and 27.5% (three-room) relative to a myopic oracle, recovering 76% and 97% of the savings a clairvoyant planner attains with the future revealed in advance. The three-room restaurant isolates the benefit of reaching beyond a single task: a one-task supervised baseline recovers almost none of the saving (3.1%), whereas our multi-task value captures most of it (27.5%).