# polar bears

- **Event:** [Gemini 4 Argon SF Hackathon](https://cerebralvalley.ai/e/gemini-4-argon-sf)
- **When:** Sun, Oct 25 at 9:00 AM – 10:00 PM (PDT)
- **Where:** San Francisco, CA
- **Placement:** 1st Place
- **Team:** [Jonathan Yang Tsang](https://cerebralvalley.ai/u/jt26), [Ben Xie](https://cerebralvalley.ai/u/benx), [Dhilen Mistry](https://cerebralvalley.ai/u/dhilen)
- **GitHub:** https://github.com/benjaminxie1/argon-hacks
- **Demo video:** https://youtu.be/dGreES9C6a0
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/67

TRY THE DEMO LIVE! labtwin.vercel.app
note: the hosted demo replays a cached run. non-demo data requires requires a local server and API key: please see the README or contact us!

LabTwin is a digital twin world model for the wet lab: Gemini 4 Argon and a swarm of Managed Agents turn a biolab protocol and one photo of your bench into a physics-based 3D digital twin, run every step before a real pipette moves, and fix the mistake before anyone spends a single sample.

 You give it three things: a protocol (PDF or Markdown), a photo or short video of your lab's tools/layout, and the result you want in each well. Gemini 4 Argon reads the protocol and the photo together and writes a spec of your bench. An orchestrator then fans out Antigravity Managed Agents in parallel sandboxes, and they write a physics twin of your bench: a whole-room MuJoCo scene, the liquid state, the reward and a Gymnasium RL environment. Argon renders the twin from where your photo was taken, critiques it against the photo and fixes itself; in our demo the visual match went from 54 to 88.

The twin then runs your protocol step by step. It compares every well’s volume and concentration with your target and traces each failed well back to the exact step that caused it. Say “Argon, fix it”: Argon rewrites the protocol with a reason for each change, and the simulator re-runs it from scratch to prove the fix. 

 Every twin is also a real RL environment: an agent trained from scratch reaches a 1.0 success rate. You can also play the protocol yourself in 3D, with mistakes flagged as you click.

## More from Gemini 4 Argon SF Hackathon

- [THRU](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/64)
- [Aniket's Team](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/65)
- [My Little Brain](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/66)
- [Footage-to-World](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/68)
- [FirmReverse](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/69)
- [SatMission](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/70)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/67. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt