# Ares

- **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:** [Chimaobi Edeh](https://cerebralvalley.ai/u/kaosi)
- **GitHub:** https://github.com/godfatherc08/Praxis
- **Demo video:** https://youtu.be/qM-T_N88eqo
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1363

PRAXIS is a self-authoring, self-verifying robot agent. It attacks the two problems today’s robot policies depend on and the open world never provides: a hand-engineered library of motor skills, and a reward signal that tells the robot when it succeeded.

PRAXIS starts from a single primitive, act: move one step in a direction, open or close the gripper. That is its entire vocabulary. Across a curriculum of manipulation tasks in LIBERO (robosuite/MuJoCo), it writes its own higher-level skills as real executable Python, runs them, and then judges the result itself from the camera pixels alone, using muse-wonder-1.0 multimodal ability as its method of self-assessment. The Meta reasoning model, muse-wonder-1.0,  is the active author of the vocabulary, proposing and selecting skills; the agent grows a persistent library it carries from task to task.

LIBERO’s ground-truth success predicate is deliberately held out from the agent and used only to grade how reliably its verifier/self-judgment matches reality. On the LIBERO-OBJECT suite, PRAXIS’s verifier made its calls with zero false positives across 17 judged attempts; it never claimed a success that wasn't, and skills authored for one task transferred untouched to five new tasks, solved on the first attempt. All of this is training-free: no demonstrations, no gradient updates, one CPU and a 8 dollars of API.

## More from Meta Global AI Developer Hackathon

- [Cascade Development Group](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1360)
- [workalike](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1361)
- [Questory](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1362)
- [Orbit Fitting Room](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1364)
- [RepoLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1365)
- [Glint](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1366)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1363. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt