What it does

RLzoo is a Python toolkit bundling common reinforcement learning algorithms (DQN, PPO, SAC, TD3, etc.) with a uniform interface over TensorFlow 2.0 and TensorLayer. You pick an algorithm name, an environment name, and a one-liner like alg.learn(env=env, mode='train') does the rest. It ships wrappers for OpenAI Gym, DeepMind Control Suite, and RLBench out of the box.

The interesting bit

The project deliberately splits into high-level APIs here and low-level tutorials in the main TensorLayer repo. That two-track design is unusual: most libraries pick one abstraction level and defend it to the death. RLzoo essentially admits that beginners and researchers need different things, and builds both.

Key highlights

• Supports implicit or explicit configuration styles: hide the knobs in default.py, or expose every network layer and optimizer in your runner script.
• Environments covered: Atari, Box2D, MuJoCo, classic control, robotics, DeepMind Control Suite, and RLBench (V-Rep/PyRep based).
• Distributed training for DPPO via Kungfu (added in v1.0.4).
• Tied to a Springer textbook on deep RL; free PDF available through institutional access.
• Paper accepted at ACM Multimedia 2021 Open Source Software Competition.

Caveats

• Default hyperparameters are explicitly noted as “may not be optimal”; benchmark results with tuned configs are promised for a future release.
• Several dependencies (MuJoCo, V-Rep, RLBench) require manual installation outside pip.
• The README warns to expect issues in the months after initial release.

Verdict

Good fit if you want to spin up standard RL experiments fast without hand-rolling training loops, or if you’re working through the companion textbook. Skip if you need battle-tested, production-hardened defaults or if you already have strong opinions about your network architectures and don’t want a wrapper layer in the way.