Apppublic

ace3848w34u32y/grid-world-navigator

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

Grid World Navigator — OpenEnv Mini RL Environment

A Mini Reinforcement Learning environment where an agent navigates a grid world to reach a goal while avoiding walls. Built for the Meta PyTorch OpenEnv Hackathon x Scaler School of Technology — Round 1.

🎯 Tasks

1.Navigate from a random start position to the goal (bottom-right corner)
2.Avoid wall collisions (walls are randomly placed)
3.Minimize steps taken (shortest path is rewarded)

🏆 Graders

Grader	Description	Range
`success`	Binary: 1.0 if goal reached, 0.0 otherwise	0.0 – 1.0
`path_efficiency`	Optimal distance / actual steps	0.0 – 1.0
`collision_score`	1.0 = no collisions, decreases with hits	0.0 – 1.0

💰 Reward Logic

Event	Reward
Each step	-0.01 (encourage shortest path)
Hit wall	-0.1 (penalize collisions)
Hit edge	-0.05 (out of bounds)
Reach goal	+1.0 (episode ends)
Max steps exceeded	-0.5 (episode ends)

🏗️ Architecture

Follows the OpenEnv 3-component pattern:

grid-world-navigator/
├── models.py              # Pydantic models: GridAction, GridObservation, GridState
├── client.py              # WebSocket client communication
├── server/
│   ├── environment.py     # Core logic: reset(), step(), state(), graders
│   ├── app.py             # FastAPI routing
│   └── Dockerfile         # Containerization
├── openenv.yaml           # Environment manifest
├── pyproject.toml         # Python package config
└── requirements.txt       # Dependencies

🚀 Quick Start

Local Development

bash

cd grid-world-navigator
uv sync && uv run server
# Or with uvicorn directly:
uvicorn server.app:app --host 0.0.0.0 --port 8000 --reload

Test health:

bash

curl http://localhost:8000/health

Docker

bash

docker build -t grid-world-navigator:latest -f server/Dockerfile .
docker run -d -p 8000:8000 grid-world-navigator:latest

Deploy to Hugging Face

bash

openenv push --repo-id username/grid-world-navigator

🎮 Action Space

Action ID	Direction
0	UP
1	DOWN
2	LEFT
3	RIGHT

📐 Observation Space

The agent receives a GridObservation containing:

—grid: 5x5 2D list (0=empty, 1=wall, 2=goal, 3=agent)
—agent_pos: [row, col] of the agent
—goal_pos: [row, col] of the goal
—steps_taken: number of steps so far
—message: feedback string

📋 API Methods

Method	Returns	Description
`reset()`	`GridObservation`	Start new episode with random layout
`step(action)`	`GridObservation`	Move agent, return updated state
`state()`	`GridState`	Full internal state for debugging

🛠️ Tech Stack

—Python 3.11+
—OpenEnv — Standardized RL environment framework
—Pydantic — Type-safe data models
—FastAPI + Uvicorn — Async WebSocket server
—Docker — Containerization

📜 License

MIT — Built for the Meta PyTorch OpenEnv Hackathon 2026.