Network architecture enforced symmetry

Deep reinforcement learning has only been recently used to control characters with continuous action space and high degrees of freedom.The locomotion produced by policies that are generated from the direct application of deep reinforcement learning are seldom symmetric.As such a direct application does not provide a mechanism to impart symmetry.One of the ways to generate more human-like locomotion is to introduce a reward to mimic motion-capture data and the other way is to introduce symmetry constraints.This thesis proposes a novel method that is hard-constrained to be state-action space symmetric and enforces temporal-state symmetry while penalizing for energy usage.This mechanism can generate a variety of natural looking gait patterns that capture the correct level of symmetry for the given task in an automated fashion.The proposed method is robust to learning a variety of desired gaits even on rugged terrain.This level of diversity is greater than an existing state-of-the-art work that uses deep reinforcement learning for symmetric locomotion.In quantitative comparisons, our method generates policies that are comparable in visual appeal and show lower variance in evaluation metrics compared to the state-of-the-art.i List of Tables 4.1 Body Dimensions for symmetric bipedal walker . . . . . . . . . . . . . . .4.2 Joint position, limit and max-torque for symmetric bipedal character . . . .4.3 Body Dimensions for asymmetric bipedal character . . . . . . . . . . . . .4.4 Joint position, limit and max torque for asymmetric bipedal character . . .4.5 Gains for reward-components corresponding to different gaits . . . . . . .5.1 Evaluation metrics of policies for different gaits generated using HACTS, on flat terrain. . . . . . . . . . . . . . .

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.