Gemini 4 Argon builds its own physics sim from one human video, lands 7 cornhole shots on Apollo
Argon reportedly watched a single human cornhole video, wrote its own physics simulator, trained a throwing policy and transferred it to the Apollo humanoid for 7 hits in a row, cutting weeks of robotics work to hours.
ONE HUMAN VIDEO. 7 SHOTS IN A ROW. Google's Gemini 4 Argon watched a person play cornhole once, built its own physics simulator, trained the throw, then moved it onto the Apollo humanoid the idea: the model isn't the robot. it's the engineer the robot part is not the news. robots have thrown things for years. the news is who did the engineering ▹ what actually happened 1. INPUT: 1 human video. no teleop sessions, no motion-capture suit, no hand-labeled dataset 2. SIM: the model wrote its own physics simulator for the bag, the board and the arm. nobody handed it one 3. TRAIN: throws get trained inside that sim, where failing costs zero 4. TRANSFER: the policy moves onto real hardware and gets refined there, closing the sim-to-real gap 5. RESULT: Apollo lands 7 shots in a row 6. TIME: Google says weeks of robotics engineering → hours. alone each step exists already. one model running all 4 in a loop is the new part VIDEO → SIM → TRAIN → REAL ROBOT → REFINE cornhole is a toy task. the loop is not. swap the bag for a box, a tool, a tray, and the bottleneck stops being robotics engineers and starts being human footage bookmark this before every humanoid demo starts with "we showed it one video"
This post quotes another post: x.com/rimtoln/status/2110306196854882586