# Mollet

- **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:** [JOOYOUNG LIM](https://cerebralvalley.ai/u/MMPORONG)
- **GitHub:** https://github.com/mmporong/mollet-meta
- **Demo video:** https://cdn.jsdelivr.net/gh/mmporong/mollet-meta@1b71e2fb2e30b3bad1f84b5e9d9e1eadf9c924f9/submission/meta-english-demo-20261105-v1/mollet-meta-demo-180s-en.mp4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3494

Mollet helps robotics developers test a recorded behavior against a requirement before another training or recovery attempt. Reward curves show change, but not when a posture or motion requirement held. Video provides context, yet watching it repeatedly does not apply the same numeric rule each time. Mollet turns the developer's natural-language intent into an editable behavior condition, aligns it with the robot's location in the video, and evaluates it against the original synchronized telemetry. The workbench reports PASS, FAIL, or UNKNOWN and shows the matching interval beside the evidence.

The interface supports condition authoring, review, editing, rerunning, approval, and JSON export. Reloading an edited export preserves its values but clears the previous approval, so a changed condition cannot inherit an old review. The evaluator supports time triggers and detects region entry when a SAM box center crosses a boundary. Tracking breaks, observation gaps, or a video too short for the required window return UNKNOWN. The authenticated H1 API example uses only a time trigger. Six recorded Isaac Lab stair-simulation clips are available across H1 biped and Go2 quadruped cases: four have synchronized time series and two are summary-only. Mollet exposes that evidence level instead of treating every clip as equally testable.

In the authenticated H1 example, the authored condition required up-axis alignment of at least 0.8 to start within eight seconds and hold for 0.5 seconds. The original telemetry returned PASS from 0.0 to 0.5 seconds. A reviewer changed the hold to eight seconds and received PASS from 0.0 to 8.0 seconds. A manually authored full-duration 12.48-second Go2 condition returned FAIL. Each result retains condition, observation, video, state, request, and response provenance through hashes, so the evaluation can be replayed without changing the source values.

Mollet builds on our pre-existing RL diagnosis foundation and recorded stair assets. The hackathon contribution is the Meta-powered condition-authoring workflow, frame location binding, deterministic evaluation, review reset, evidence export and reload, saved-response replay, and an extension for video-grounded hypotheses and reflection. The extension registers an original MP4 with synchronized state, derives one to four hypotheses that cite frames present in that source, checks up to three previously untested conditions on the same stored observation, and prepares reflection. A separately confirmed triage probe can support a registered correction proposal after case and correction checks; the proposal stays pending human approval with a GPU budget of zero. These paths passed synthetic-response CPU and browser tests, but no new video-model call, real reflection, reward change, retraining, GPU job, or measured improvement is claimed. Source: https://github.com/mmporong/mollet-meta. Demo: https://mollet-meta.vercel.app.

## More from Meta Global AI Developer Hackathon

- [23Editor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3491)
- [StepProof](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3492)
- [V&F](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3493)
- [Remboai](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3495)
- [Creative Conventionals](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3496)
- [Threadline Discovery](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3497)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3494. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt