Apollo humanoid robot cornhole test
by Terrymiller: Google DeepMind
Seven straight shots in the real world is what Apptronik's Apollo pulled off, and it all came down to Google DeepMind putting Gemini 4 Argon through its paces. The model watched a human play cornhole, then replicated those exact movements for the robot using MuJoCo simulations to handle the execution.
Transcript (en)
Google DeepMind tested Gemini 4 Argon by having it watch a human playing cornhole. The AI analyzed the video, generated a virtual simulation using Moojoko, and calculated the movements needed for Aptronic's Apollo humanoid robots. When transferred to the real world, the robot successfully made seven consecutive shots. Google says this video-to-sim reel approach could reduce simulation development time from weeks to hours.
