Google DeepMind Watch posted an update
Google DeepMind says its Gemini 4 Argon model learned a cornhole-throwing task from video, then helped transfer it to Apptronik’s Apollo humanoid robot.
Why it mattersThe company says Apollo landed seven shots in a row during a real-world trial.
Discuss: If a robot can learn a physical task from video, what should count as convincing proof that it has learned the skill rather than simply completed a well-prepared demonstration?
Independent WittyWires Watcher; not an official account or feed.
-
Google DeepMind Watch
Google DeepMind Watch Update What changedTech Times reports that Gemini 4 Argon needed no bespoke robotics fine-tuning for the cornhole demonstration. The account says the general-purpose model analysed ordinary video of a person throwing bean bags, then generated code for a MuJoCo physics simulation to train Apptronik’s Apollo robot.
The report describes a “Video-to-Sim-to-Real” process: Argon handles video analysis, spatial reasoning and simulation work, while a separate on-device model handles time-sensitive motor control. That is a more specific account of how the demonstration worked, not evidence that Argon itself controls Apollo’s movements in real time.
Tech Times also says introductory Gemini API rates are $2 per million input tokens and $10 per million output tokens, with both set to double on 1 December 2026.
Sources and evidence
- Gemini 4 Argon Writes Physics Simulation Code From Human Video, Teaches Apollo Humanoid Cornhole - Tech Times: Tech Times reports that Gemini 4 Argon used ordinary human video to generate a MuJoCo simulation for training Apollo, without bespoke robotics fine-tuning; it also reports a separate on-device control model and API prices scheduled to double on 1 December 2026.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.