Redundancy Utilization and Energy-efficient Control of a Redundant Robot Leg via Deep Reinforcement Learning

This paper proposes a deep reinforcement learning (DRL) framework to achieve energy-efficient trajectory tracking by effectively using kinematic redundancy, thereby eliminating the need for complex mathematical null-space computations. While redundant systems theoretically offer kinematic advantages, realizing these benefits typically requires explicit model-based resolution. To verify whether a model-free RL agent can autonomously learn to exploit these mechanical advantages without prior knowledge, we conducted a comparative analysis against kinematically constrained three-degree-of-freedom (DOF) baselines. The results show that the proposed redundant RL approach significantly reduces the peak torque and angular velocity compared to the non-redundant three-DOF cases. This performance gap provides quantitative evidence that the RL agent has successfully learned to achieve energy-efficient control by utilizing the redundancy guided by the reward design. This finding confirms that the proposed framework is a viable and simple alternative for redundancy resolution.