Reflex: Reinforcement Learning with Reflection Symmetry Exploitation in State-Based Continuous Control

1. Dependencies

This project follows the CleanRL style and uses modules from the CleanRL ecosystem.

• CleanRL repository: https://github.com/vwxyzjn/cleanrl
• CleanRL documentation: https://docs.cleanrl.dev/

Please follow the instructions in cleanrl to install dependencies.

Notes:

• If you run dm_control/* tasks, make sure shimmy[dm-control] is installed.

2. Run Script

Note: CPU-only is recommended for faster and more stable experiment, since most runtime is spent on environment interaction in state-based RL rather than policy optimization. We used an Mac M4 (10 CPU cores, 32GB RAM) in practice.

run_exp.py is the experiment launcher. By default, it uses multiprocessing to train multiple random seeds in parallel.

Run experiments with:

python run_exp.py -c configs/baseline_sac.yaml
python run_exp.py -c configs/reflex_sac.yaml

Use different config files to run different experiment settings.

Single-process mode (one seed)

You can disable multiprocessing and run with a single process:

# Use the first seed listed in the config file
python run_exp.py -c configs/baseline_sac.yaml --single-process

run_exp.py reads these fields from the selected config file:

• model: training script path
• environments: environment list
• seeds: random seed list
• total_timesteps: per-environment training steps

So selecting a different file in configs/ will run a different training script and experiment setup.