Planning with Q-Values in Sparse Reward Reinforcement Learning

Learning a policy from sparse rewards is a main challenge in reinforcement learning (RL). The best solutions to this challenge have been via sample inefficient model-free RL algorithms. Model-based RL algorithms are known to be sample efficient but few of them can solve sparse settings. To address these limitations, we present PlanQ, a sample efficient model-based RL framework that resolves sparse reward settings. PlanQ leverages Q-values that encode long-term values and serve as a richer feedback signal to actions than immediate rewards. As such, PlanQ scores rollout returns from its learned model with returns containing Q-values. We verify the efficacy of the approach on robot manipulation tasks whose difficulties range from simple to complex. Our experimental results show that PlanQ enhances performance and efficiency in sparse reward settings.