Although remarkably common in the animal world, two-legged locomotion has been a particularly challenging open research problem. The ultimate goal of bipedal walking research is to develop a framework that enables an agent to learn to ‘walk’ in any unknown environment where this is permitted by the laws of physics. This implementation seeks to train an agent to maintain human gaits with maximum horizontal velocity in the Mujoco Humanoid-env. A first-principles approach to the implementation of the Soft Actor-Critic (SAC) stochastic policy optimization algorithm has been discussed and it’s slow-convergence and potential implications analyzed.

Please refer Project_Report.pdf for further details about the algorithm and it's implementation details.

Please check out the original paper at Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor
You may also find OpenAI's tutorials on this to be very helpful.