# navfail

- **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:** [Manjunath Jayaramaiah](https://cerebralvalley.ai/u/manju89jay)
- **GitHub:** https://github.com/manju89jay/navfail
- **Demo video:** https://www.youtube.com/watch?v=GcGkKBwSlXA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2448

When a mobile robot fails to navigate (stops in front of an open gap, never arrives), an engineer has to find which of hundreds of navigation settings caused it, by hand: read the logs, guess, rerun the simulation, repeat. Every guess costs a simulator run, and a lucky run can look like a fix.

navfail does that investigation. Built during the hackathon, it has three parts:
1. A reproduction harness that rebuilds a recorded Nav2 failure in Gazebo on demand and records every run, including the robot's true trajectory.
2. An investigating agent (Muse Wonder 1.0 via the Meta Model API) that reads the same evidence an engineer gets: the operator's note, the filtered Nav2 log, a timeline extracted from the rosbag2 recording, and the deployed configuration. It proposes one change as strict JSON, only among allowlisted settings.
3. A frozen referee that tests each proposal (one screening run, then 4 more plus 3 on an empty floor, judged on the true position from the simulator) and recomputes every reported number from the raw run files. Rejected proposals and their results go back to the agent.

On a planted failure we raced the agent against random search, 5 searches each with a 40-run budget. The agent proved its first proposal every time, in 8 runs (the minimum); random search needed 8 to 28 (median 8 vs 9). The whole race cost $0.46 of API. Every number can be re-checked from a fresh clone with verify.py; the limits are stated in the README.

## More from Meta Global AI Developer Hackathon

- [Twimler](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2445)
- [Merlin](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2446)
- [La Asistenta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2447)
- [Relight](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2449)
- [One Market](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2450)
- [IntentLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2451)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2448. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt