Heterogeneous computing is a promising approach for supporting emerging parallel applications with diverse performance, accuracy, and energy demands. Effective task scheduling in such systems should optimize multiple competing objectives, including energy efficiency, computational performance, and accuracy, while accounting for task dependencies and communication overhead. In this paper, we propose a reinforcement learning (RL)-based multi-objective scheduler designed to improve the energy efficiency of highly heterogeneous computing systems. Our approach adapts scheduling decisions to minimize energy consumption while maintaining accuracy and performance.