# Kintsugi

- **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:** [Yahya Khalid](https://cerebralvalley.ai/u/yahyak)
- **GitHub:** https://github.com/0xE1G/sensei
- **Demo video:** https://www.youtube.com/watch?v=abXK6yvgmlE
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4104

Kintsugi is a robot studio where robots build new skills themselves and prove they work. It's built around reliability and efficiency and runs on a normal laptop CPU with no training or GPU needed.

THE PROBLEM: Nowadays, robots learn skills from thousands of demonstrations and days of GPU training. And when a failure is found, the loop repeats. Training RL policy also requires substantial compute and designing reward functions is difficult.

HOW IT WORKS: You describe a task in plain language like "open the drawer", "stack blocks from largest to smallest" or show your real workspace. Muse Spark plans the task, and from a photo, SAM 3.1 and Muse Image rebuild your workspace as a digital twin. A robot arm then performs the task in simulation using skills already in library and a plan from Muse Spark. When a step fails or a new skill is needed that isn't in the library yet, it diagnoses the failure and deliberately generates more failure scenarios, and then fixes it by tuning a skill or creating a completely new one with Meta's Muse Spark reasoning capabilities. Every diagnosis is tested: a cause is only accepted after replaying exact scene with one change and then verifying it.

Furthermore, it takes reliability seriously and keeps a fix only when it proves with statistical tests that it works on scenes and objects it has never seen and doesn't break anything that worked before. The engine is built in such a way that it reasons over geometry rather than blindly letting an LLM write code. Skills are defined on shapes the cameras measure from their perception, like a rim, a top face or a long axis, and the LLM never writes a trusted skill directly: simulation finds the numbers and the tests decide if a skill should be kept.

Another advantage of this system is that it is very efficient. When a skill is already present, it transfers zero-shot to unseen objects of similar shapes without needing an LLM to write any skill, just a plan. For example, it acquired a grasping skill on a bowl, and when it saw a mug for the first time, it could pick it up zero-shot because it had already picked up a similar shape (a rim). Checking it on the mug took 10 trials instead of 900 to build the skill from scratch. This makes the system efficient, reduces LLM calls and makes it more reliable. In short, the system is reliable, general, efficient and, most importantly, very cost-efficient. In my testing, it acquired completely new skills for about 1 cent on average and tasks with skills already present cost almost nothing.

RESULTS: 475/480 scenes succeed across 10 tasks like sorting, stacking, pushing and opening drawers. Muse Spark wrote 4 missing skills (push, stand-up, open drawer, close drawer) and all passed the tests on first try.

Everything happens in one studio: you plan, run, investigate, repair and re-run in one web app with live 3D, replay every episode, see why failures happened, and every skill is versioned with its evidence and history written by Spark.

## More from Meta Global AI Developer Hackathon

- [SwayamVGoswami](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4101)
- [Yonro](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4102)
- [Solo Anchor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4103)
- [CrowdMap](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4105)
- [BARA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4106)
- [HotBot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4107)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4104. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt