# Armed

- **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:** [Pranay Joshi](https://cerebralvalley.ai/u/pranayjoshi)
- **GitHub:** https://github.com/pranayjoshi/armed-metaAI-hackathon
- **Demo video:** https://youtu.be/r3KbxNEi8mc
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5116

Armed is a third arm that autocompletes you.

Here is the problem it goes after. For users who rely on switch scanning to operate a helper robot, the routine is: wait for the highlight to land on the thing you want, hit the switch, wait again. A CHI 2022 study (Bonaker et al.) measured about five selections a minute at more than two switch hits each. That is the state of the art for the people who need a robot arm most. And for everyone else, a robot arm is something you have to stop and instruct, which is slower than just doing the thing yourself.

So I flipped it. Armed reads the clock, the room, who is speaking and one muscle in your forearm, and proposes what you are about to need. The phone lights up while you type: “Bring the phone?” One twitch, and it is at eye level. Water time, no water on the desk: it brings your shake to the straw. You hold out a jar on a call and it offers to hold it while the lid comes off. Someone else asks it to pass the marker, and it waits for you before it moves. Every yes all day is one muscle. Words spoken to it: zero.

The part I am proudest of is what happens when it gets something wrong. It carried the drink straight over my laptop. I tensed my forearm and it froze mid-air, with no model anywhere in that loop. Wonder investigated the clip and joint traces, identifying the crossing before exhausting its tool budget. I authored the restriction from that finding; my checker validated it against the recorded incident, and the next carry went around the laptop. I changed a robot’s future behaviour without saying a word or opening a menu.

Four Meta models do the seeing, hearing, judging and investigating. Everything that moves is my own code: taught poses, a pure-function gate, a stop with no model or network dependency. Model proposals must pass my permission gate before anything moves. Nine days, one person, about $700 of open hardware, and the first two of those days went on assembling the arm, calibrating it and getting the first picks to repeat. I think this is how every helper robot should be built, and I built one.

## More from Meta Global AI Developer Hackathon

- [Not Random](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5113)
- [Ares-Crucible](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5114)
- [Bucketwise | Karthik Ramani](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5115)
- [Forge Code](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5117)
- [MetaCare](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5118)
- [Children Of Exu](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5119)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5116. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt