README.md

---
license: mit
task_categories:
- reinforcement-learning
tags:
- robotics
- mujoco
- continuous-control
---

# ARC-RL: A Reinforcement Learning Playground Inspired by ARC Raiders

This repository contains the expert demonstration datasets for **ARC-RL**, a suite of four MuJoCo continuous-control environments featuring robotic morphologies inspired by the game *ARC Raiders*.

- **Paper:** [ARC-RL: A Reinforcement Learning Playground Inspired by ARC Raiders](https://huggingface.co/papers/2605.19503)
- **GitHub Repository:** [https://github.com/CarloRomeo427/ARC_RL](https://github.com/CarloRomeo427/ARC_RL)

## Robot Morphologies
The benchmark features four unique robotic body plans:
- 👑 **Queen**: 18-DoF hexapod.
- 🛡️ **Bastion**: 12-DoF armoured hexapod.
- 🐸 **Leaper**: 12-DoF quadruped.
- 🪲 **Tick**: 18-DoF compact hexapod.

The datasets provided here are hand-crafted Central Pattern Generator (CPG) demonstrations for each morphology, which serve as fixed expert references and as sources of prior data for offline-to-online reinforcement learning algorithms like SACfD, SPEQ-O2O, and SOPE.

## Usage

To launch training with prior data (which automatically downloads the necessary files from this repository), use an algorithm like `sacfd` or `sope`:

```bash
python main.py --algo sacfd --env queen --log-wandb
```

To regenerate or collect expert episodes manually:

```bash
python -m src.utils.collect_dataset --env bastion --n-episodes 1000
```

## Citation

```bibtex
@misc{romeo2026arcrlreinforcementlearningplayground,
      title={ARC-RL: A Reinforcement Learning Playground Inspired by ARC Raiders}, 
      author={Carlo Romeo and Andrew D. Bagdanov},
      year={2026},
      eprint={2605.19503},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.19503}, 
}
```