Home · Technology · Oct 23 archive
Google DeepMind's Gemini 4 Argon Teaches Robot New Skills from Video
Confirmed
In Short: Google DeepMind demonstrated its Gemini 4 Argon AI model teaching Apptronik’s Apollo humanoid robot new physical skills by watching videos.
Google DeepMind has demonstrated a new robotics training technique using its Gemini 4 Argon AI model. The model learned a physical task by watching video and then transferred the resulting behavior to Apptronik’s Apollo humanoid robot.
In a real-world trial, Apollo successfully landed seven consecutive shots after being trained on a video of a person playing cornhole. Google DeepMind says the 'Video-to-Sim-to-Real' pipeline can compress weeks of manual simulation modeling into a few hours.
Gemini 4 Argon autonomously generated simulation code before the learned behavior was transferred to the robot. The model achieved an 87 percent score on its Embodied Reasoning benchmark, which measures its ability to reason about physical and spatial problems in 3D environments.
The demonstration shows how general-purpose AI models could reduce the amount of manual engineering required to teach robots new physical skills. Google DeepMind describes this approach as a 'Video-to-Sim-to-Real' pipeline, transforming ordinary videos into a scalable, infinite training ground for robotics.
Gemini 4 Argon is designed for complex engineering, scientific, coding, and business tasks. The model combines visual-semantic reasoning, physical understanding, and autonomous code generation for robotics applications.
In an internal test, Argon was shown raw video of a person playing cornhole. It then mapped the arm swing and release point onto a humanoid’s joints, built a full physics simulation of the robot and the board, practiced the motion, and transferred the result to Apptronik’s Apollo humanoid.
Background
Google has launched Gemini 4 Argon, a new AI model designed for complex, multi-step tasks, available to developers and enterprises via API and to subscribers of Google AI Ultra and Pro.
Google has launched Gemini 4 Argon, a new AI model designed for complex, multi-step tasks, available to developers and enterprises via API and to subscribers of Google AI Ultra and Pro.
What's confirmed
- Google DeepMind has demonstrated a new robotics training technique using its Gemini 4 Argon AI model. The model learned a physical task by watching video and then transferred the resulting behavior to Apptronik’s Apollo humanoid robot.
- In a real-world trial, Apollo successfully landed seven consecutive shots after being trained on a video of a person playing cornhole. Google DeepMind says the 'Video-to-Sim-to-Real' pipeline can compress weeks of manual simulation modeling into a few hours.
- Gemini 4 Argon autonomously generated simulation code before the learned behavior was transferred to the robot. The model achieved an 87 percent score on its Embodied Reasoning benchmark, which measures its ability to reason about physical and spatial problems in 3D environments.
- The demonstration shows how general-purpose AI models could reduce the amount of manual engineering required to teach robots new physical skills. Google DeepMind describes this approach as a 'Video-to-Sim-to-Real' pipeline, transforming ordinary videos into a scalable, infinite training ground for robotics.
- Gemini 4 Argon is designed for complex engineering, scientific, coding, and business tasks. The model combines visual-semantic reasoning, physical understanding, and autonomous code generation for robotics applications.
- In an internal test, Argon was shown raw video of a person playing cornhole. It then mapped the arm swing and release point onto a humanoid’s joints, built a full physics simulation of the robot and the board, practiced the motion, and transferred the result to Apptronik’s Apollo humanoid.
What's still developing
- Nothing material beyond the confirmed record.