Figure 5.

Impact of the symmetry coefficient (w_s) on the learned motor control based on (A) accuracy, (B) agility, (C) metabolic efficiency, (D) range of motion, and (E) symmetry. For each sub-figure, six models were independently trained with different w_s parameters and their performances were quantified. As shown, bigger w_s coefficient incentivizes the agent to learn more symmetric activation patterns. However, such model is less accurate, and is not able to explore as much of the state space. Moreover, the non-symmetric penalty (Equation 10) has a regularization effect which decreases the overall neural excitations and increases the metabolic efficiency (ME). Accordingly, the model trained with w_s = 10 (circled in all sub-plots) demonstrates a balanced performance across the metrics.