Investigating Object Permanence in Deep Reinforcement Learning Agents

Abstract

Object Permanence (OP) is the understanding that objects con- tinue to exist when not directly observable. To date, this ability has proven difficult to build into AI systems, with Deep Rein- forcement Learning (DRL) systems performing significantly worse than human children. Here, DRL Agents, PPO and Dreamer-v3 were tested against a number of comparators (Hu- man children, random agents and hard coded Heuristic agents) on three object permanence tasks (OP) and a range of con- trol tasks. As expected, the children performed well across all tasks, while performance of the DRL agents was mixed. Over- all the pattern of performance across OP and control tasks did not suggest that any agent tested except children showed evi- dence of robust OP.