# Eupho

- **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:** [Harsha Vardhan](https://cerebralvalley.ai/u/harshadots), [Sejal Kishor Daterao](https://cerebralvalley.ai/u/SKDaterao)
- **GitHub:** https://github.com/harsha-iiiv/MIRROR
- **Demo video:** https://youtu.be/PXZg3Z-cA4g
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4068

MIRROR is a robot task-authoring and testing studio built on MuJoCo 3.17. Today, teaching a robot a task means writing code or a reward function, and the robot's understanding stays invisible until it fails. MIRROR flips that: demonstrate the task, inspect what the system thinks matters, and correct a misunderstanding in plain language.

The problem it solves: a robot can pass its specification and still do the wrong thing. We filmed the task on a real table with one phone. MIRROR rejected three calibration clips before one passed, tracked the mug, red box, marker, tray and divider with Muse Spark and SAM 3.1, and asked 8 questions. Its camera check found the zoom estimate 6.6% off against a 6% limit, so it refused to infer the task and ran the stated task instead: mug, then marker, then red box, into the tray around the divider (simulated: success, 0 violations, 8/8 beads kept). We corrected it in one sentence: "Keep the mug at least 6 cm away from the divider while carrying it." The first run, re-judged from its step log, violated the new rule (closest 4.44 cm); the re-planned run kept 9.44 cm and satisfied all rules. That take is a development take, never scored; Meta model spend, all real footage: about $1.10.

A correction in MIRROR becomes one scoped rule change (one rule, on named objects, in a named phase), checkable three ways: old runs are re-judged from execution evidence, with unsettled rules reported as "not verified", never as a pass; the corrected rule is planned and run on layouts it was never written for; and when no compliant route exists, MIRROR refuses with a proof instead of breaking the rule. On 10 blocked layouts where the only way through tilts the cup, the original specification completes all 10 runs, spilling 38 beads at each of six rotation prices; the corrected specification plans nothing and returns a proven infeasibility record on all 10.

Two pre-registered evaluations on fresh layouts, planned at a rotation price of 0.05 m/rad: violation-free runs went from 27/60 and 33/60 to 60/60 (and 4/30 and 3/30 to 30/30 on a shortcut family), with beads lost falling 119 to 0, 106 to 0. Caveat, published with the numbers: with the price calibrated to execution time (0.2445 m/rad), the uncorrected plan never tilted the cup on 120 new layouts with an upright route, so the cup rule changed nothing there. Ordering and clearance corrections do change behaviour at the calibrated price on development layouts: "Pack the marker before the cup" is violated by old runs on 6/6 and changed on 6/6; "Keep the cup at least 6 cm away from the wall" is violated on 6/6 (old runs pass 3.0-4.1 cm) and satisfied on 6/6 (new runs keep 8.0-9.1 cm).

Simulation only: MuJoCo physics, a free-flying parallel-jaw gripper plus an optional UR10e arm, no physical robot, no sim-to-real claim. Every verdict comes from an independent checker reading the simulator's step log, never from a model's description. Live studio: https://mirror-robot-studio.netlify.app

## More from Meta Global AI Developer Hackathon

- [ModStorm](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4065)
- [MAVIS (Solo)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4066)
- [Bottega](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4067)
- [xMemory Hunters](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4069)
- [Shedrack](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4070)
- [Reacknadam](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4071)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4068. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt