This repository provides the official implementation of the paper:

Actor-Free Continuous Control via Structurally Maximizable Q-Functions
Yigit Korkmaz*, Urvi Bhuwania*, Ayush Jain†, Erdem Bıyık†
NeurIPS 2025 (to appear)

We introduce Q3C, a framework for continuous control without an explicit actor network by leveraging structurally maximizable Q-functions for stable and efficient policy learning.

conda create --name q3c_env python=3.10

conda activate q3c_env

pip install -r requirements.txt
pip install -e .

Note: MuJoCo must be installed separately, as it is required for all benchmark environments.

This repository uses Hydra for hierarchical configuration management.
The file configs/hydra_sb3_q3c_hyperparams.yaml defines environment-specific hyperparameters.

To train and log results to Weights & Biases, run:

python scripts/run_experiment.py wandb.entity=YOUR_WANDB_USERNAME train.environment=ENV_NAME

• Replace YOUR_WANDB_USERNAME with your W&B account name.
• Replace ENV_NAME with your target environment (e.g., Hopper-v4, HalfCheetah-v4, Walker2d-v4).

All metrics, losses, and rollout statistics will be automatically logged to your W&B project.

If you find this repository useful for your research, please cite:

@inproceedings{korkmaz2025actor,
  title={Actor-Free Continuous Control via Structurally Maximizable Q-Functions},
  author={Korkmaz, Yigit and Bhuwania, Urvi and Jain, Ayush and Bıyık, Erdem},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2025}
}