title	My Env Environment Server
emoji	💿
colorFrom	yellow
colorTo	green
sdk	docker
pinned	false
app_port	8000
base_path	/web
tags	openenv

Maintenance-Bench

An OpenEnv-based reinforcement learning environment for industrial maintenance diagnostics.

This repository provides the initial implementation of a diagnostic decision-making environment designed to evaluate AI research agents on industrial troubleshooting tasks.

The environment is inspired by real-world industrial maintenance logs and is part of the Maintenance-Bench benchmark.

---

Overview

The environment simulates an industrial diagnostic process where an agent must decide which diagnostic steps to perform to identify a fault.

Each episode corresponds to a maintenance case scenario stored as a JSON file.

A scenario contains:

• machine metadata
• symptoms
• possible diagnostic steps
• observations for each step
• rewards
• operational costs

The agent interacts with the environment through a step-by-step diagnostic process, receiving observations, rewards, accumulated costs, and the current health status of the system.

---

Environment API

The environment follows the OpenEnv interface:

• reset() – loads a new diagnostic scenario and initializes an episode
• step(action) – performs a diagnostic action and returns the observation

---

Episode Flow

1. The environment loads a diagnostic scenario from the data/ directory.
2. The agent receives the initial symptom description.
3. The agent chooses a diagnostic action.
4. The environment returns:

• the observation
• reward
• accumulated cost
• health status of the system
• history of performed steps

---

Example Interaction

Example sequence of actions:

reset()

Agent observation:

"High Bearing temperature."

Agent performs diagnostic step:

action = "diagnose:0"

Environment response:

Observation:
"Overload relay found tripped at the time of breakdown."

Reward: 1
Total Cost: 5
done: False #equivalent to health status 0

History:
[
  {
    "action": "Monitored bearing temperatures at DE and NDE using RTD.",
    "observation": "Overload relay found tripped at the time of breakdown."
  }
]

---

Project Goal

The goal of this project is to provide a benchmark environment for evaluating deep research agents on industrial maintenance diagnostics tasks.

This environment supports research in:

• autonomous troubleshooting agents
• cost-aware reasoning
• industrial decision making
• long-horizon diagnostic planning

---

---

Running the Environment (OpenEnv + Docker)

The environment is built using OpenEnv, which automatically packages the environment into a Docker image that exposes an HTTP API.

Build the Environment

OpenEnv builds the environment and creates a Docker image:

openenv build

This process:

• packages the environment
• creates a Docker image
• prepares the HTTP server interface

---

Run the Environment

After building, run the environment container:

docker run -p 8000:8000 my_env_env:latest

This starts the environment server and exposes it at:

http://localhost:8000

---

Available Endpoints

Once running, the following endpoints are available:

POST /reset
POST /step
GET  /state
GET  /health

You can interact with the environment using:

• OpenEnv clients
• HTTP requests
• Swagger UI
• custom research agents

License

This project builds on OpenEnv and follows the license terms of the respective components.