# Aurora

- **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:** [Divyansh Agrawal](https://cerebralvalley.ai/u/divagr18)
- **GitHub:** https://github.com/divagr18/nemesis
- **Demo video:** https://www.youtube.com/watch?v=JirwnL-4Ed8
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2723

Nemesis is a real-to-sim reinforcement learning platform for robotics, built for the Meta AI Hackathon. It uses Meta's Muse models and SAM 3.1 to turn natural language, voice, and images of real environments into physics simulations where legged robots can be stress-tested and retrained with PPO.

The platform lets you break and improve robot policies without writing simulation code. Describe a difficult terrain, apply external forces, or upload a real-world scene. Nemesis recreates the conditions in MuJoCo, tests the robot, and lets you retrain it directly on its failures.

It also includes several autonomous and interactive modes:

- Autopilot: Muse agents independently search for failure modes and investigate weaknesses.
- You vs Muse: Compete against Muse to find the smallest realistic perturbation that knocks a robot over.
- Playground: Interact with a live robot simulation in your browser, push it around, and train it to recover.
- Live Deployment: Describe a deployment environment and have Muse generate, simulate, and evaluate realistic scenarios.
- Readiness Reports: Inspect policy robustness, failure modes, and deployment limits.
- Humanoid Support: Run the same stress tests on a Unitree G1 humanoid alongside the Go1 quadruped.

Every retrained policy passes a regression gate that checks improvement on the targeted failure, normal locomotion, and performance across 52 fixed test scenarios before deployment.

In evaluations, a single round of adversarial retraining reduced falls on held-out scenarios from 27.5% to 13.5%, without degrading normal walking performance. Nemesis also supports a fully self-hosted training and evaluation workflow on a single 24 GB GPU.

Try it live: https://nemesis.divagr.com

## More from Meta Global AI Developer Hackathon

- [ActCheck](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2720)
- [Violet Sky Security](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2721)
- [Quack](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2722)
- [Prasanna Sharma](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2724)
- [DAFE Trading Systems](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2725)
- [Sabbk (Nudgling)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2726)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2723. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt