Google DeepMind has demonstrated a robotics training technique in which its Gemini 4 Argon model learns a physical task by watching video, then transfers the resulting behavior to Apptronik's Apollo humanoid. The work forms part of the launch of Gemini 4 Argon, which Google says combines visual-semantic reasoning, physical understanding and autonomous code generation for robotics.
Key takeaways
- In an internal test, Argon was shown raw video of a person playing cornhole; the model analyzed the environment, mapped spatial coordinates and calculated the whole-body joint movements needed for the task.
- Argon then autonomously generated simulation code, and the learned behavior was transferred to Apollo, which landed seven consecutive shots in a real-world trial.
- Google DeepMind calls the approach a "Video-to-Sim-to-Real" pipeline and says it compresses weeks of manual simulation modeling into a few hours.
- On the company's Embodied Reasoning benchmark (ERQA), which tests reasoning about physical and spatial problems in 3D environments, Argon scored 87 percent.
- Gemini 4 Argon is now available through Google's AI products and the Gemini API; Google has also applied it to large codebases, quantum computing, nuclear fusion and life-sciences research.
For manufacturing automation, the demonstration points to a shorter path from demonstration footage to deployable robot behavior, potentially reducing the manual engineering needed to commission new physical tasks on humanoid platforms. The result is a single internal trial, and broader validation across industrial tasks remains to be shown.
Source: Robotics & Automation News · Published 2026-10-22T19:34:44 · “Google DeepMind uses Gemini 4 Argon to teach Apptronik’s Apollo humanoid new skills from video”