AI Learns to Walk, But Less Dumb

A developer shares progress on an AI locomotion project inspired by attending Nvidia's GTC conference. The post covers building a graphical editor for physical model assembly, fixing mass/density issues, and most importantly switching from evolutionary algorithms to PPO (Proximal Policy Optimization) reinforcement learning using the MLAC C++ library. The comparison between the two approaches is striking: evolutionary methods plateau quickly and produce inelegant 'vibrating' solutions, while PPO produces smooth, physically convincing locomotion after about 30 minutes of training. The project uses a pogo-stick-like structure with collider constraints to force proper balancing behavior.