---
license: mit
library_name: stable-baselines3
pipeline_tag: reinforcement-learning
tags:
- reinforcement-learning
- stable-baselines3
- soft-actor-critic
- sac
- robotics
- continuous-control
- gymnasium
- mujoco
- reacher-v5
---


# SAC Agent for Reacher-v5 🎯

This repository contains a trained **Soft Actor-Critic** agent for the Gymnasium `Reacher-v5` environment. 

The agent controls a two-joint robotic manipulator and was trained using Stable-Baselines3, Gymnasium and MuJoCo.

[View the complete project on GitHub](https://github.com/M-Stasiak/TSwR_project)


## Usage

Clone the complete project from GitHub and install the required dependencies:

```bash
git clone https://github.com/M-Stasiak/TSwR_project.git
cd TSwR_project
pip install -r requirements.txt
```

### Option 1: Manual download

[Download the trained SAC model](sac_reacher_final.zip) and place it in the `trained_models` directory.

### Option 2: Download with `huggingface_hub`

Install the Hugging Face client:

```bash
pip install huggingface_hub
```

Download the model directly into the required directory:

```python
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="M-Stasiak/sac-reacher-v5",
    filename="sac_reacher_final.zip",
    local_dir="trained_models",
)

print(f"Model downloaded to: {model_path}")
```

Then select the RL controller in `main.py` and run:

```bash
python main.py
```

The GitHub repository contains the custom environment wrapper, observation processing and simulation code required to run the trained model.


## Model details

- **Algorithm:** Soft Actor-Critic
- **Environment:** `Reacher-v5`
- **Framework:** Stable-Baselines3
- **Policy:** `MlpPolicy`
- **Network architecture:** `[256, 256, 256]`
- **Training timesteps:** `2,000,000`
- **Observation size:** `10`
- **Action size:** `2`


## Evaluation results

The final evaluation was performed over 1000 episodes.

| Metric                  | Result |
| ----------------------- | -----: |
| Success rate            |  97.2% |
| Average number of steps | 48.341 |


## 📜 License

This model is released under the **MIT License**. See the [LICENSE](LICENSE) file for details.