# Locus

- **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:** [Madhur Thareja](https://cerebralvalley.ai/u/madhurthareja), [Anushka Pawar](https://cerebralvalley.ai/u/A1431)
- **GitHub:** https://github.com/madhurthareja/MetaLocus
- **Demo video:** https://youtu.be/uFx60F0nBs0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5105

Locus : spatial working memory for assistive navigation.

Most AI wearable available to a visually impaired person today is a stateless captioner. Ask it what's ahead and it answers well. Turn your head, and its world empties. The bollard, let's say, it described two seconds ago no longer exists to it, while you keep walking toward it.

For sighted users this is a minor flaw, because a sighted person's head points where they're looking. A visually impaired user doesn't scan. The camera points wherever the head happens to be. Field-of-view loss isn't an edge case for this population, it's the normal operating condition. That makes persistent memory a precondition for head-mounted assistive vision to work.

Locus adds the missing layer. Muse Spark 1.3 reads a keyframe twice a second and decides what matters in this scene, the open-world vocabulary, since you cannot enumerate street hazards in advance. SAM 3.1 grounds those words to pixels; EdgeTAM tracks them causally at 10.5 FPS. Masks are lifted to metric 3D through dense stereo depth and written into a world-frame graph that keeps objects alive after they leave the camera. Geometry, range, bearing, time-to-collision, closest point of approach, decides what is worth saying, and SeamlessM4T v2 says it.

Memory turns out to do two jobs, it answers where something is after you've looked away: 0% out-of-view answer rate for a stateless baseline, 24% for a sliding-window baseline, 100% with the graph. And it is why the system can stay quiet, a stateless system has no identity across frames, so it cannot know the pole it sees now is the pole it saw two seconds ago, and cannot tell a stationary pole from one apparently moving at 13 m/s. Fixing that took alerts from 342 of 539 frames down to a single advisory event.

Assistive devices fail by being annoying, not by being wrong. Locus is built to know when not to speak.

## More from Meta Global AI Developer Hackathon

- [nah i'd lose](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5102)
- [Open CMO Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5103)
- [Translator](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5104)
- [Hygge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5106)
- [AURA RLVE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5107)
- [Slate HQ Ltd](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5108)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5105. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt