tukaram12/RL-task-manager

Grid-based RL Mini-Game Environment

Project Description

This is a comprehensive, production-ready reinforcement learning environment designed for Hugging Face Spaces. It strictly follows OpenEnv API standards. The game challenges an agent to navigate a 10x10 grid from a starting point (top-left) to a goal point (bottom-right) while avoiding traps and collecting optional rewards.

API Endpoints (FastAPI)

The environment provides the standard RL endpoints via FastAPI:

• —POST /reset Returns: {"observation": <Observation>}
• —POST /step Expects Body: {"direction": "<up|down|left|right>"} Returns: {"observation": <Observation>, "reward": <float>, "done": <bool>}
• —GET /state Returns: <Observation>

Observation and Action Spaces

Observation (Pydantic Model)

• —agent_position: Tuple[int, int]
• —goal_position: Tuple[int, int]
• —grid_size: int
• —steps_taken: int

Action (Pydantic Model)

• —direction: Literal string containing either "up", "down", "left", or "right".

Reward Logic

• —Base turn penalty: -1.0 (encourages taking the shortest path)
• —Reaching the goal: +10.0
• —Hitting a trap: -5.0
• —Collecting a reward cell: +3.0

Graders & Difficulties

The environment specifies three difficulty tasks:

• —Easy: Just reach the goal.
• —Medium: Reach the goal efficiently with fewer steps.
• —Hard: Reach the goal under a strict, unforgiving step-limit.

Instructions to Run Locally

1. 1.Install dependencies:

   pip install -r requirements.txt

1. 1.Start the OpenEnv FastAPI Server:

   uvicorn app:app --host 127.0.0.1 --port 7860

1. 1.Run the Streamlit UI (in a new terminal window):

   streamlit run streamlit_app.py

1. 1.Evaluate the baseline agent:

   python inference.py

Deploying to Hugging Face Spaces

This project includes a standard Dockerfile setup for Docker runtime SDK on Hugging Face Spaces. Simply push the repository to a space with SDK set to Docker, and the API will be exposed on port 7860 automatically.