Combining On-Policy and Off-Policy RL to Improve Bipedal Walker: A Tutorial on Hybrid Reinforcement Learning Approaches

This preprint presents preliminary results from an ongoing research effort investigating a hybrid reinforcement learning pipeline combining Proximal Policy Optimization (PPO), behavior cloning, and Deep Deterministic Policy Gradient (DDPG) to improve gait stability in the BipedalWalker environment. The presented experiments are intended as a proof of concept. Further evaluations involving multiple random seeds, further ablation studies, and alternative off-policy algorithms are planned.