This is the code base for my bachelor thesis on bipedal locomotion using reinforcment learning.
See https://scripties.uba.uva.nl/search?id=record_28025 for the full thesis.
Legged robots are far more mobile than their wheeled counterparts, but designing a controller for legged locomotion poses many challenges. Terrain-blind locomotion controllers are highly robust to uneven terrain as they do not rely on detailed sensing of the terrain. While recent RL techniques have shown great capacity to learn, direct application does not yield a robust controller. We explore the applicability of the method from Lee et al. to bipedal locomotion using the HardcoreBipedalWalker environment. Multiple agents are trained using different approaches and compared, and insight into the use and applicability of the method is presented.