Learning Efficient Representations for Goal-conditioned Reinforcement Learning via Tabu Search

An intelligent robot should understand the user-specified goals and execute corresponding skills to fulfill them. In this paper, we study the problem of visual goal-conditioned robot manipulation. In contrast to previous methods, our method improves the successful rate of visual goal-conditioned tasks and the generation ability toward different goals from the aspect of network structure. We also use Tabu Search to select higher-parameters for best learning results. To that end, we propose a slot-based network within Goal-Conditioned Reinforcement-Learning framework. The network can focus on task-related areas while filtering out environmental disturbance. We trained and validated the effectiveness of the proposed network in two simulation environment tasks. The experiments proved that the proposed slot-based network can efficiently extract goal-conditioned information and successfully complete the given task compared to baselines.