Conservative Offline RL with Uncertainty-Aware Policy Improvement

Implement an offline RL pipeline based on Conservative Q-Learning (CQL) or IQL with uncertainty-aware policy improvement. Use Gymnasium datasets (e.g., D4RL or synthetic logged data) for tasks such as HalfCheetah, Hopper, and Walker2d; if MuJoCo is unavailable, fall back to classic control tasks with logged data. Compare against baselines (BC, TD3+BC). Report normalized scores, variance across seeds, and ablations on conservatism weight.