ichi-goat7/cloud-kitchen-env

🍳 Cloud Kitchen Scheduling Environment (OpenEnv)

🚀 Overview

This project implements a real-world reinforcement learning environment simulating a cloud kitchen operation.

An AI agent must intelligently schedule incoming food orders to limited cooking slots (stoves), while managing:

• —⏱️ Cooking time
• —📦 Order deadlines
• —🔥 Limited kitchen resources
• —💰 Revenue optimization

The goal is to maximize total reward by completing high-value orders on time while avoiding inefficiencies.

🧠 Environment Design

The environment follows a standard OpenEnv-style API implemented as a FastAPI app.

Core Endpoints

• —POST /reset → Initializes the environment and returns the initial observation
• —POST /step → Executes an agent action and returns observation, reward, done, info
• —GET /state → Returns the current environment state
• —GET / → Health check (returns "Cloud Kitchen Env running 🚀")

Core Methods (Backend)

• —reset() → Initializes environment
• —step(action) → Executes agent action
• —state() → Returns current observation

🔗 Base URL

https://ichi-goat7-cloud-kitchen-env.hf.space

🔄 1. Reset Environment

Endpoint

POST /reset

Description

Resets the environment to the initial state.

Request

No body required.

Example (PowerShell)

Invoke-RestMethod -Uri "https://ichi-goat7-cloud-kitchen-env.hf.space/reset" -Method POST

Example (curl)

curl -X POST https://ichi-goat7-cloud-kitchen-env.hf.space/reset

Response

{ "current_time": 0, "orders": [...], "slots": [...] }

⚡ 2. Take a Step

Endpoint

POST /step

Description

Executes one environment step using the provided action.

Request Payload

{ "assignments": [ { "slotid": 1, "orderid": 1 } ] }

Example (PowerShell)

Invoke-RestMethod -Uri "https://ichi-goat7-cloud-kitchen-env.hf.space/step" -Method POST -Body '{"assignments":[{"slot_id":1,"order_id":1}]}' -ContentType "application/json"

Example (curl)

curl -X POST https://ichi-goat7-cloud-kitchen-env.hf.space/step \ -H "Content-Type: application/json" \ -d '{"assignments":[{"slotid":1,"orderid":1}]}'

Response

{ "observation": { "current_time": 1, "orders": [...], "slots": [...] }, "reward": 0.0, "done": false, "info": { "events": [] } }

📊 3. Get Current State

Endpoint

GET /state

Description

Returns the current environment state without modifying it.

Example (PowerShell)

Invoke-RestMethod ` -Uri "https://ichi-goat7-cloud-kitchen-env.hf.space/state"

Example (curl)

curl https://ichi-goat7-cloud-kitchen-env.hf.space/state

Response

{ "current_time": 0, "orders": [...], "slots": [...] }

✅ Action Format Summary

{ "assignments": [ { "slotid": int, "orderid": int } ] }

🏁 Example Workflow

1. 1.POST /reset
2. 2.POST /step (assign orders)
3. 3.Repeat /step until "done": true

📥 Observation Space

Each step returns:

• —current_time → Current timestep
• —orders → List of pending orders
• —slots → Current cooking slot status

🎮 Action Space

The agent provides assignments of order_id → slot_id

Example

Action(assignments=[
    Assignment(slot_id=1, order_id=2)
])

🔄 Design Note

Initially, the environment supported explicit manual assignment of orders to slots.

In the current version, the baseline agent automates this process using a scheduling strategy:

• —When a slot becomes free, the agent automatically selects an order
• —Orders are chosen based on Earliest Deadline First (EDF)

👉 This reflects a more realistic system where decisions are made dynamically rather than manually.

🎯 Motivation

Unlike toy problems or games, this environment models a real operational challenge faced by cloud kitchens:

"How do you prioritize and schedule orders when resources are limited and deadlines matter?"

This introduces real-world complexities like:

• —Trade-offs between speed vs value
• —Resource contention (limited slots)
• —Time-sensitive decision making

🏆 Reward Function

The reward system reflects operational efficiency:

✅ Positive Rewards

• —Full reward → Order completed before deadline
• —Partial reward → Order completed slightly late

❌ Penalties

• —Late completion → Reduced or negative reward
• —Idle cooking slot → Penalty (unused resource)

👉 This encourages the agent to keep slots busy and prioritize timely deliveries

📊 Tasks & Difficulty Levels

We define 3 tasks with increasing difficulty:

Interpretation

• —Agent performs well in simple settings
• —Performance drops under constraints
• —Demonstrates increasing task complexity

✅ Key Features

• —✔ Real-world task (cloud kitchen scheduling)
• —✔ OpenEnv-compatible API
• —✔ Typed models using Pydantic
• —✔ Multi-task evaluation system
• —✔ Continuous reward shaping
• —✔ Reproducible baseline results
• —✔ Dockerized for deployment

⚙️ Setup Instructions

1. Clone Repository

git clone https://github.com/KumoDrift/CLoudKitchenEnv-pytorch-hackathon
cd cloud-kitchen-env

2. Install Dependencies

pip install -r requirements.txt

▶️ Run Inference

python inference.py

🐳 Docker Support

Build and run:

docker build -t cloud-kitchen-env .
docker run -p 7860:7860 cloud-kitchen-env

👨‍💻 Author

KumoDrift Cloud Kitchen RL Environment Project