Smart Factory Scheduling

A reinforcement learning environment that models the scheduling problem as it actually appears on a factory floor: jobs arrive with deadlines and priorities, machines break down stochastically, and the agent has to keep throughput up without missing due dates.

It implements the OpenEnv interface, so it drops into any training loop that already speaks that protocol.

Observation space. Machines carry an id, a status (idle, busy, broken), the job they are currently running, and a failure rate. Pending jobs carry remaining time, deadline, priority on a 1–3 scale, and their assigned machine. The agent also sees completed jobs, the current timestep, and the episode horizon.

Why this problem. Job-shop scheduling is a clean testbed for decision-making under stochastic disruption. The interesting behaviour isn’t optimal assignment on a static job list — it’s what a policy does when a machine dies mid-job and the schedule it was committed to no longer exists.

«FILL: Results. What did you train against it, and what happened? Even “PPO beats the greedy earliest-deadline-first baseline on throughput but loses on deadline misses” is a real finding. If you haven’t trained anything on it yet, say that plainly — an honest “environment built, baselines next” is fine and better than silence.»

The repository is packaged as a Docker-based Hugging Face Space.

«TODO: deploy the Space and link it here. The front matter in your README already has sdk: docker set up for it. A clickable demo is worth more than this entire write-up.»