Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Reinforcement Learning Learning Track (Gymnasium + SB3 + SOTA Libraries)

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.

Suggested Order

  1. 01_cartpole_dqn_sb3_basics
  2. 02_lunarlander_ppo_a2c
  3. 03_bipedalwalker_sac_td3
  4. 04_atari_breakout_dqn_cleanrl
  5. 05_mujoco_halfcheetah_rl_zoo_tuning
  6. 06_fetchpush_her_goal_conditioned
  7. 07_minigrid_recurrentppo_curriculum
  8. 08_pettingzoo_multiagent_parameter_sharing
  9. 09_offline_rl_minari_d3rlpy
  10. 10_imitation_learning_bc_gail

Project Map

  • 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).

Core Libraries Used Across Projects

Environment Setup (uv)

  1. cd /Users/user/Projects/rl-envs
  2. uv lock (refreshes uv.lock when dependencies change)
  3. uv sync (creates/updates .venv and installs from uv.lock)
  4. Run scripts with uv run, for example:
    • uv run python 01_cartpole_dqn_sb3_basics/train.py
    • uv run python 01_cartpole_dqn_sb3_basics/eval.py --model-path 01_cartpole_dqn_sb3_basics/artifacts/best_model/best_model

Recommended General Study References

Practical Note

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.

About

A collection of lightweight, modular environments designed for training and benchmarking reinforcement learning agents. This repository focuses on providing clear, accessible implementations of classic and custom MDP (Markov Decision Process) challenges.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages