Your desk, rehearsed by robots before hardware. A phone video becomes a physics twin, Gemini 4 Argon critiques each failed hand-over, and MuJoCo decides what passes.
Film a table with your phone. ArgonTwin turns it into a physics twin, and two ALOHA robot arms try a hand-over on it in MuJoCo. When an attempt fails, Gemini 4 Argon reads the rendered frames and the measured numbers, explains why and patches the plan. MuJoCo runs it again and decides.
The model proposes. Physics decides.
Results
The demo run: the arm hit the screens at 216 N and the can fell off the table. Argon’s patch passed on the next attempt, 5 mm from target with zero contact.
Five live repairs, all fail to pass. Each Argon call reads about 27,000 tokens of frames and metrics.
A fine-tuned Gemma 4 E4B, running locally, fixed 11 of 12 held-out failures, up from 4 of 12 for the base model.
The twin: a 10.4-second iPhone video, 52 camera poses, a Gaussian splat trained in 10 seconds on a Colab T4.
