# Say-and-Grasp

- **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:** [Jonathan Adoo](https://cerebralvalley.ai/u/adoodevv)
- **GitHub:** https://github.com/adoodevv/say-and-grasp
- **Demo video:** https://youtu.be/6_HE9kJhO6w
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/472

Say-and-Grasp turns a $250 SO-101 robot arm into Wiggle, a voice-controlled third hand for the workbench. Say "Wiggle, pass me the screwdriver" and it finds the tool, picks it up and drops it into your open hand, with zero task-specific training. Say "tidy up" and it plans and puts every loose tool into a tray. Ask "where's my multimeter?" and it points to it or says what it last did with it. It has a little personality too: it wakes up when you call its name, waves, nods, dances and nods off when idle.

The problem: makers, repair technicians and electronics labs constantly stop work to reach for tools, and for people who can't easily reach across a bench it's a real barrier. Low-cost arms usually need dozens of human demonstrations for every task. Wiggle needs none.

It also builds its own training data as it works. Every fetch and put-away is recorded as a LeRobot episode (both cameras, joints, actions, task text) and labelled success or fail by Muse Spark. Wiggle collected 52 episodes (51 successful, zero human demonstrations), published openly on Hugging Face, and I trained a first ACT policy on them with Hugging Face Jobs. It reaches toward the block but doesn't grasp yet; more episodes are the next step, and collecting them is exactly what the loop does. The zero-shot system is the teacher.

Results: 10/10 fetches into the hand (median 22 s from request to palm), 83% of tools put away in multi-tool tidy-up runs.

Built during the hackathon: everything in the repository. Existing before: the SO-101 arm hardware and its LeRobot motor calibration.

## More from Meta Global AI Developer Hackathon

- [geometry os](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/469)
- [solo q](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/470)
- [RescueLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/471)
- [SIRE — Smart Ingestion & Reconciliation Engine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/473)
- [AegisVerity Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/474)
- [VERITAS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/475)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/472. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt