This workspace now contains 10 project folders arranged from beginner to advanced.
Each project has a detailed README.md with goals, milestones, evaluation criteria, and helper docs/blogs.
01_cartpole_dqn_sb3_basics02_lunarlander_ppo_a2c03_bipedalwalker_sac_td304_atari_breakout_dqn_cleanrl05_mujoco_halfcheetah_rl_zoo_tuning06_fetchpush_her_goal_conditioned07_minigrid_recurrentppo_curriculum08_pettingzoo_multiagent_parameter_sharing09_offline_rl_minari_d3rlpy10_imitation_learning_bc_gail
- 01: Discrete control fundamentals with DQN (CartPole).
- 02: Policy gradient methods on Box2D (LunarLander) with PPO/A2C.
- 03: Continuous control with SAC/TD3 (BipedalWalker).
- 04: Pixel observations and Atari preprocessing (Breakout).
- 05: MuJoCo benchmarking + hyperparameter tuning (HalfCheetah).
- 06: Sparse-reward robotics via HER + goal-conditioned RL (FetchPush).
- 07: Partial observability + memory/curriculum in MiniGrid.
- 08: Multi-agent RL with PettingZoo and scaling with RLlib.
- 09: Offline RL using Minari datasets and d3rlpy (CQL/IQL).
- 10: Imitation learning pipelines (BC + DAgger).
- Gymnasium
- Stable-Baselines3
- SB3 Contrib
- Gymnasium-Robotics
- MiniGrid
- PettingZoo
- Ray RLlib
- Minari
- d3rlpy
- PyTorch
cd /Users/user/Projects/rl-envsuv lock(refreshesuv.lockwhen dependencies change)uv sync(creates/updates.venvand installs fromuv.lock)- Run scripts with
uv run, for example:uv run python 01_cartpole_dqn_sb3_basics/train.pyuv run python 01_cartpole_dqn_sb3_basics/eval.py --model-path 01_cartpole_dqn_sb3_basics/artifacts/best_model/best_model
- Gymnasium Basic Usage
- SB3 RL Tips and Tricks
- Spinning Up in Deep RL
- Hugging Face Deep RL Course
- David Silver RL Course
Use the single project environment managed by uv for all folders and log experiments with TensorBoard or Weights & Biases.
For projects 8-10, expect longer training runs and higher compute requirements.