FULL STORY

Gemini 4 Argon: Google's Physical Reasoning Model Makes Waves

Google DeepMind launched Gemini 4 Argon, a general physical reasoning model opened to all users, wowing observers by teaching a robot cornhole from video alone. Follow-up official blog posts and developer experiments showing paper replication in a single prompt kept the momentum going.

2026-10-13 ~ 2026-10-14 · 4 episodes · 15 posts

Episode 1 · Google DeepMind's Gemini 4 Argon Learns Cornhole from Video, Hits Seven in a Row on Apollo, Now Open to All (2026-10-13, 6 posts)

Google DeepMind has released and fully opened Gemini 4 Argon, and a robot demo has drawn wide attention: after merely watching videos of humans playing cornhole, the model analyzed the physics on its own, built a virtual simulation in MuJoCo, wrote its own practice code, and transferred what it learned to Apptronik's Apollo humanoid robot—hitting seven out of seven throws. The model is now available to all users, marking a landmark demonstration for physical AI.

Confirmed

  • Google DeepMind has repeatedly confirmed the demo details: after watching videos of humans playing cornhole, Gemini 4 Argon autonomously mapped the environment, computed full-body joint movements, and automatically set up a virtual environment in MuJoCo simulation for practice.
  • The company officially launched Gemini 4 Argon for Robotics, positioned as a general physical reasoning model, scoring 87% on pure reasoning on the 3D physical-space understanding benchmark Embodied Reasoning (ERQA).
  • Officials announced that Gemini 4 Argon is now available to all users, saying it can turn video into a training ground for physical AI: in the cornhole demo, a modeling process that previously took weeks was compressed to a few hours.
  • The robot hardware in the demo is Apptronik's Apollo humanoid (referred to as Apollo 2 in some posts), achieving seven consecutive hits.

Why it matters

  • The demo showcases a complete loop of "video understanding + autonomous simulation + embodied transfer": without hand-written simulations or step-by-step demonstrations, the model completes the entire process from observing physics to generating a training environment on its own.
  • If the efficiency gains highlighted by officials (weeks of modeling down to hours) can be replicated across more tasks, it would significantly lower the barrier for robot skill learning—a major step toward general physical AI.

Episode 2 · Google says Gemini 4 Argon is a step change for robotics (2026-10-14, 2 posts)

Google published a developer blog showcasing how Gemini 4 Argon accelerates robotics research, with DeepMind's Pushmeet Kohli and others calling it a "step change," including the ability to directly control quadruped and humanoid robot whole bodies.

Episode 3 · Gemini Argon reproduces robotics papers with a single prompt, simulates environments from video (2026-10-14, 2 posts)

keerthanpg demonstrated that Gemini 4 Argon can reproduce multiple robotics papers with a single prompt and zero-shot generate simulation environments from real first-person videos, covering both human-to-robot and robot-to-simulation paths.

Episode 4 · Google Launches Gemini 4 Argon for Robotics, Hitting 87% on ERQA (2026-10-14, 5 posts)

Google introduced the robotics foundation model Gemini 4 Argon for Robotics on its official developer blog, positioning it as a general-purpose physical reasoning model and calling it a major step toward "physical AGI."

Confirmed

  • The model combines robot control, physical reasoning, simulation environment generation, and skill improvement into one integrated capability (@DynamicWebPaige)
  • It achieved 87% through pure reasoning on Embodied Reasoning (ERQA), a benchmark for embodied reasoning that tests problem solving in 3D physical spaces, outperforming all other frontier models (@IMordatch, @keerthanpg)
  • Simulation training results: 4,500 simulation training runs boosted the success rate from 56% to 75%; the model can build digital simulations from real-world data, test actions in virtual environments first, and then deploy them to physical robots (@DynamicWebPaige)
  • Zero-shot manipulation: it can control different types of robots zero-shot; a demo video shows Boston Dynamics' Spot quadruped pausing to "think" in simulation when facing complex tasks, with the reasoning process playable as voice (@bousmalis)
  • Another demo: it taught a humanoid robot to play cornhole zero-shot just by watching video; it can also reproduce robotics papers from a single prompt (@bousmalis, @keerthanpg)

Why it matters

The 87% ERQA score shows the model surpasses other frontier models in 3D physical reasoning, and its workflow of building simulations from real data, testing virtually, then deploying physically—along with zero-shot cross-robot control—could, if validated, significantly lower the barrier to developing robot skills, marking major progress on the path toward general-purpose robotics foundation models.