# PiSpace Innovations

- **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:** [Abebe Sendek](https://cerebralvalley.ai/u/airo7)
- **GitHub:** https://github.com/airo7/airo-meta
- **Demo video:** https://youtu.be/kepW0BCZU-c
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2616

Robots break quietly when their bodies change. Every robot drives by a model of its own body (how far and how fast a motor command moves it), tuned once by an engineer. Then it carries a heavy tote, tows a cart, its tyres wear or its battery sags, and the model is wrong: the robot stops short or drifts and doesn't know it. Today the fix is a service call. Over a robot's life, setup, tuning and upkeep cost more than the robot itself (purchase is 25–40% of lifetime cost), and 98% of manufacturers are small businesses with no robotics engineer on staff.

AIRO (Adaptive Intelligence for Robots) lets a robot notice that its body changed, test itself and recalibrate in minutes, with one phone camera and no wheel encoders. Muse Wonder acts as the scientist: it designs each test drive and states a falsifiable prediction before the wheels turn. The robot measures the result, fits a physics body model with calibrated uncertainty, and only accepts a new body once a prediction it made lands. On our real $50 car with a water bottle strapped on, the old model predicted 0 cm and the car moved 17; after relearning, 7 predictions in a row landed. Every learned body goes into a memory the whole fleet shares: in the MuJoCo warehouse twin, the first robot loaded with a heavy tote learns it in about 12 s, and every later robot recognises it in 2.1 s with one probe (41/41 heavy loads identified, 0 false relearns, 180/180 deliveries over 10 stress runs). Because the robot knows its body, you can talk to it: Muse Spark turns plain words into a plan, Muse Voice Transcribe hears it, SAM 3.1 reads the floor, and Muse Image draws each body's self-portrait.

## More from Meta Global AI Developer Hackathon

- [hannibal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2613)
- [Quint Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2614)
- [Donkers](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2615)
- [Ron Bee Software](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2617)
- [Velvora Nexus](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2618)
- [futurpreneur](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2619)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2616. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt