# Liam O'Neill

- **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:** [Liam O'Neill](https://cerebralvalley.ai/u/LiamO)
- **GitHub:** https://github.com/Liamostudio/Pathfinder
- **Demo video:** https://youtu.be/MCpxhh_a9II
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4084

Pathfinder is a live safety system for robot arms. The idea is simple: anyone should be able to put a robot arm on a bench they share with people, and work next to it safely, without an industrial budget.

Robot arms have got cheap. Mine is a 3D-printed SO-101 with about £150 of servos and a power supply. But keeping one safe around people still usually means cages, light curtains or keeping everyone out of the way. That's fine for factories, not for a workshop, a classroom or a maker space. So I built the safety side using things most people already have: two webcams, a printed marker and a laptop.

How it works:

- Map the workspace. A few photos of the bench become a 3D map, using Meta's MapAnything for the geometry and SAM 3.1 to pick out the table, walls and robot. A printed marker ties it all to the real world.
- Mark where the arm can't go. The bench becomes a grid of cubes, and you mark no-go areas. The arm treats those as hard rules and goes over or around them, or tells you why it can't.
- Watch the people. Two webcams track hands and people. A fast hand check runs locally, so the arm slows as your hand gets close, stops near it and carries on once you move away. SAM 3.1 adds people to the 3D map about once a second. The fast, simple check does the stopping; the slower model fills in the picture.
- Just tell it what you want. I talk to it through Meta AI glasses: "pick up the red cube and put it next to the blue one". Muse Voice Transcribe turns that into text and Muse Spark works out the steps. Spark can only use five tools (look at the scene, pick, place, stop, say), and every move is checked before it happens. If something gets in the way it waits, retries or finds another way, and if it really can't finish it tells you why and goes back to rest.
- The models never drive the motors. They suggest; a small local safety layer decides. None of the safety depends on the internet, and saying "stop" overrides everything.

What's real: the proximity stop in the video is the real arm, which I built and calibrated during the hackathon. It slows, holds still about 10 cm from my hand, carries on when I move away, then folds back. The picking and placing is a simulated version of the same arm, running through the same hand check, models and safety rules. The cameras, hand tracking, SAM 3.1, Voice Transcribe and Muse Spark are all running live.

Everything was built during the hackathon, starting straight after kickoff on 26 October. PROGRESS.md in the repo is a dated log of what was built and tested, and the README is upfront about what's real, what's simulated and what isn't finished yet.

## More from Meta Global AI Developer Hackathon

- [I tried my best](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4081)
- [Utkarsh Shukla](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4082)
- [Walkable](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4083)
- [art-sparrow](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4085)
- [SebastianMejia](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4086)
- [X-Arc](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4087)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4084. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt