# TwinBench

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [JeongWook Kim](https://cerebralvalley.ai/u/Pixaxe)
- **GitHub:** https://github.com/pixaxe0826/twinbench-irm1
- **Demo video:** https://www.youtube.com/watch?v=v6p5AqdUmvA&t=2s
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2230

TwinBench is a physics-driven digital twin built to test AI maintenance agents safely before they touch a real machine. It models IRM-1, an inspection and reject module on a packaging line, generated entirely from a single spec file: conveyor, guided pneumatic reject cylinder, valves, sensors, control panel, and PLC logic. Motion is computed, not animated. A pneumatic model sets the cylinder stroke, a drive model sets the belt speed, and the PLC times out and retracts a weak cylinder, so a leak produces a short stroke and a missed reject.

Eight faults from a catalogue (air leak, regulator drift, shifted reed switch, open solenoid, dirty photo-eye, belt slip, motor overload, load-cell drift) are injected into the physics. A diagnostic agent must find the root cause, isolate every energy source with lockout/tagout, repair the fault, and pass a hidden twin check that confirms the machine runs again. A repair without verified isolation is blocked and counted as unsafe.

The project includes an interactive 3D web app where you can run the line, inject a fault, and press “Run agent” to watch the agent diagnose and repair the live twin step by step. In 48 real evaluation episodes, the agent with a twin re-check after repair succeeded 79.2% of the time versus 66.7% without it (not yet statistically significant), made 6 versus 23 wrong repair attempts, and never took an unsafe action. We also found and fixed problems in our own evaluation, including a tool response that leaked the answer, and re-ran everything.

## More from Meta Global AI Developer Hackathon

- [Lukas Campos](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2227)
- [doudou](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2228)
- [BillRadar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2229)
- [WhatsGene](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2231)
- [Sorted](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2232)
- [GLPfit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2233)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2230. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt