Google's Gemini 4 Argon teaches humanoid robot cornhole zero-shot from a single video, tops ERQA at 87%
bousmalis · x · 2026-10-14
Google released Gemini 4 Argon, a robotics-focused model built for physical understanding, code generation and visual-semantic reasoning, calling it a step toward physical AGI.
- Scores 87% on the Embodied Reasoning (ERQA) benchmark via pure reasoning, beating all other frontier models
- Flagship demo: given only a video of a human playing cornhole, the model zero-shot taught an Apptronik Apollo humanoid — it took control, generated its own sim environment, performed sysID, trained on the real robot, then landed consecutive airmails after a few real shots and explained its own process
- Four core use cases: robot control (MuJoCo digital twins, zero-shot control, closed-loop steering of video world models), Real2Sim (calibrated twins from camera feeds, egocentric video, or paper sketches), physical understanding (dexterity labeling, dense value functions), and research/self-improvement on hardware
- DeepMind's bousmalis: this will be "seismic for robotics"
Related event: Google Launches Gemini 4 Argon for Robotics, Nears Physical AGI(4 posts)→
More from Embodied
- Robot uses MuJoCo as a tool for kinematics and motion planning, succeeds in 6 real-world trials — ki_ki_ki1 · 2026-10-14
- Startup Squad Turns Idea Into Robot-Drawn Brooklyn Bridge Sketch in 30 Minutes — davidfromkansas · 2026-10-14
- Stream Ring, an AI wearable three years in the making, starts shipping next month — pzakin · 2026-10-14
- Dyna-2 Robots Predict the Future: Million-Hour World Models and the Bitter Lesson of Robotics — JasonMa2020 · 2026-10-14
- AI2's MolmoAct 2 Tops Human-Preference Robot Rankings; 55K-Vote Dataset Open-Sourced — allen_ai · 2026-10-14
- Google's Gemini 4 Argon for robotics: simulation training lifts task success from 56% to 75% — DynamicWebPaige · 2026-10-14