Teaching a Robot to Walk — RL Project with 5 Algorithms & Live Demo

Built a complete end-to-end reinforcement learning pipeline to train a virtual bipedal robot to walk, then benchmarked 5 AI algorithms across 3 training strategies to find what actually drives performance.

What was done:

- Trained 5 RL algorithms from scratch: PPO, SAC, TD3, A2C, TQC

- Designed a custom adaptive curriculum system that automatically increases terrain difficulty as the robot improves

- Performed full statistical analysis with 95% bootstrap confidence intervals

- Deployed a live interactive demo on HuggingFace Spaces

Key result: TQC achieved 320 mean reward with 100% success rate. Algorithm choice explained 2.65x more variance in performance than training strategy — proven consistently across all 3 trials.