Back to articles

Google Introduces Gemini 4 Argon for Robotics and Simulation

3 min read

Introduction

Google has introduced Gemini 4 Argon as a general-purpose physical reasoning model, rather than a policy network dedicated solely to producing robot actions. Its demonstrations place the model at several points in a robotics workflow: interpreting video, constructing simulated environments, producing annotations, controlling robots, and iteratively improving policies. This broader role could be useful, but it also makes the boundaries of the evaluation harder to interpret.

Key points

  • Robot control: Google says Argon guided a Boston Dynamics Spot robot through a foot-pedal task after training in a MuJoCo twin environment. In another demonstration, it controlled a simpler robotic arm using visual coordinates and HTTP motion primitives. A separate example used the Gemini Omni video world model for closed-loop control.
  • Building training assets from images and video: A Real2Sim example reconstructed a MuJoCo scene from teleoperation video. Video annotation prompts were used to label demonstrations for behavioral cloning. Google also showed a domain-randomized MuJoCo scene synthesized from a single RGB image and used to fine-tune a vision-language-action policy. The company reports that the policy’s success rate in the corresponding real environment rose from 56% to 75%.
  • Benchmark and access: Google reports that Argon scored 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using reasoning alone, without auxiliary tools, and says the result leads other frontier models on that test. Argon is available to Google AI Ultra subscribers and through the paid Gemini API under the model ID gemini-4-argon.

What the demonstrations do—and do not—show

The examples do not establish that Argon performs every robotics task without external structure. The published prompts include extensive task-specific instructions, coordinate grids, motion primitives, and validators. These are part of the demonstrated system, not incidental details. Google has reported individual VLA examples but has not published failure rates for each demonstration. The ERQA score and task-specific success figures therefore measure different things and should not be treated as interchangeable evidence of capability.

Simulation fidelity is another important qualification. A scene may look realistic while still representing contact, friction, or other physical behavior inaccurately. For manipulation, matching what happens at contact can matter as much as matching the pixels. Combining a frontier model with simulation and quality checks could reduce handoffs between separate tools, but the available material does not show that Argon can replace dedicated VLA policies, simulators, or annotation services. Further results, including failure cases and tests of simulated versus real contact behavior, will be important for judging how general the approach is.

Source: InfoQ 中文

Comments

Checking sign-in status...

Loading comments...

Related articles