# Humanoid Controller

- **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
- **Team:** [Abhijit Betigeri](https://cerebralvalley.ai/u/abhijitbetigeri)
- **GitHub:** https://github.com/abhijitbetigeri/humanoid-lift-lab
- **Demo video:** https://humanoid-lift-lab.onrender.com/assets/humanoid-lift-lab-60s.mp4
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/32

Humanoid Lift Lab: Gemini 4 Argon fixes a humanoid's lift in simulation before deployment.

Humanoids can do the lift-and-shift work of small businesses, workshops and households: boxes, stock and furniture parts, often in cramped spaces where getting extra people to help is hard. Every load is different, and a lift tuned for one box can tip the robot or drop the next. There is no robotics engineer on site to re-tune it, and testing on the real robot risks broken boxes, damaged robots or injuries. Our loop finds and fixes those failures in MuJoCo first: the robot attempts the lift, we render the attempt, Argon watches the video with numerical telemetry, diagnoses the physics and rewrites a readable Python controller. A deterministic physics scorer checks the next attempt; the model never grades itself, and a keep-best rule discards any rewrite that regresses.

Starting from an engineer-written controller that lifted 22–88 lb but dropped a light 11 lb box, both video-plus-telemetry and telemetry-only adaptation went from 7/8 to 8/8 tested loads (11–88 lb) in four revision rounds, fixing the 11 lb drop without losing the heavier lifts. Success requires at least 0.45 m of rise, a 3-second hold, no fall or drop, and mean balance assist at or below 60 N. Starting instead from a textbook stoop-lift controller, no arm reached a pass in 6–8 rounds; that is our stated limit.

The live demo takes a request by chat or voice ("lift an 11 pound box"), runs it in MuJoCo with regression checks at 33 and 66 lb, shows Argon's timestamped diagnosis and the exact code diff it made, and re-tests. A results site shows every round, critique and video, and a voice companion answers spoken questions about the saved evidence.

These are development results in a modified, assisted G1-based simulation (torques scaled to full-size humanoid class, in-place lift only), not hardware validation or held-out generalization. Both feedback conditions reached 8/8, so we do not claim video superiority.

## More from Gemini 4 Argon SF Hackathon

- [ArgonTwin](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/29)
- [JCZ](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/30)
- [babtuna](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/31)
- [celestial2](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/33)
- [Team Real2Play](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/34)
- [AI Pitch Studio](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/35)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/32. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt