This page lists key packages developed or maintained by the AutoRL community that support research in RL and its intersection with Hyperparameter Optimization (HPO), Dynamic Algorithm Configuration (DAC), and Generalization. These packages cover a broad range of functionalities—from benchmark suites and testbeds to modular RL libraries—enabling reproducible experiments, contextual adaptation, and scalable evaluation of learning algorithms.

• ARLBench: A benchmark suite for hyperparameter optimization and neural architecture search in reinforcement learning, offering fast, JAX-based implementations of DQN, PPO, and SAC across diverse environments to evaluate HPO methods efficiently.
• CARL: A context-adaptive RL benchmark library that extends standard environments with configurable physics and task parameters, enabling interpretable evaluation of agents’ generalization capabilities.
• HPORLBench: A benchmark and toolkit for hyperparameter optimization in reinforcement learning, providing static and dynamic benchmark data handlers and example scripts to reproduce performance analyses.
• DACBench: A library for Dynamic Algorithm Configuration benchmarking, focusing on reproducibility and comparability of DAC methods via surrogate benchmarks and containerized experiment setups.
• Mighty: A unified contextual reinforcement learning library integrating Gymnasium, CARL, and DACBench environments, offering modular implementations of DQN, PPO, and SAC with built‑in logging and HPO compatibility via Hypersweeper.
• MDP Playground: A Python package featuring toy MDPs and complex environment wrappers for Gym, Atari, and MuJoCo to inject and analyze dimensions of hardness, designed as a testbed for debugging and benchmarking RL agents.