Artificial Intelligence
Google pitches Gemini 4 Argon for robot tasks
Google presented Gemini 4 Argon as a single robotics model for control, simulation and training-data review, but its examples and benchmark claims also showed heavy prompting and mixed results.
- Google described Gemini 4 Argon as a single robotics model for control, simulation and training-data evaluation.
- In one demo, a Boston Dynamics Spot robot practised a task in a virtual bin before trying it on real hardware, with two AI agents checking its positions.
- Google said Argon scored 87 percent on ERQA, while Artificial Analysis's AA-Robotics placed it third with 55 out of 100.
- Google said the model was available through Google AI Studio, Antigravity and the paid Gemini API tier for Google AI Ultra subscribers.
- Google did not give failure rates for the individual demos or say how many attempts underlay the 75 percent success figure.
Google presented Gemini 4 Argon as a model for robot control, simulation and training-data review, and said it was another step towards physical AGI. In one demo, a Boston Dynamics Spot robot practised a can-toss task in a virtual bin before trying it on real hardware, with two other AI agents checking its foot and gripper positions.
The published prompt for one cup task used a centimetre grid and preset commands. Other examples turned a photo, a teleoperated robot recording, a first-person human video and a sketch into simulations or control plans.
Google said one photo produced 77 tasks and more than 4,500 training episodes, and it said real-robot success rose from 56 percent to 75 percent as execution time fell from 35 to 24 seconds. The article said Google did not state how many attempts sat behind that 75 percent figure.
Google also paired Argon with Gemini Omni, which generated video clips from step-by-step instructions while Argon checked frames and changed the next command. In that demo, five letter tiles were arranged to spell ARGON after 33 steps, and Google also showed Argon reviewing its own logs and camera images to improve a Lego-brick task.
Google said Argon scored 87 percent on ERQA, which it described as ahead of other frontier models. Artificial Analysis's AA-Robotics benchmark put it third at 55 out of 100, behind GPT-6 Astra and GPT-6.1 Sol, and Google said Argon was available in Google AI Studio and Antigravity for Google AI Ultra subscribers, and through the paid Gemini API tier.
Named in this story
Companies
- published the Argon demos and benchmark claims
- Artificial Analysis
- published the AA-Robotics benchmark
Organisations
- DeepMind
- was named as Google's AI subsidiary pursuing physical AGI
Products and systems
- Gemini 4 Argon
- was Google's robotics model for control, simulation and training-data review
- Boston Dynamics Spot
- carried out the can-toss demo
- Gemini Omni
- generated video clips from Argon's instructions
- ERQA benchmark
- measured embodied reasoning in 3D environments
- Claude Opus 5.5
- was Google's ERQA runner-up comparison model
- GPT-6 Astra
- appeared in both benchmark rankings
- Claude Fable 5.1
- was another ERQA comparison model
- AA-Robotics
- tested real robot control across eight tasks
- Google AI Studio
- was one access route for Argon
- Antigravity
- was another access route for Argon
- Google AI Ultra
- was the subscription tier required for access
- Gemini API
- offered paid access to Argon
- Gemini Robotics-ER 1.6
- was Google's earlier spatial perception and task-planning model
How the source tells it
The article was sceptical and explanatory, pairing Google's promotional language with prompt-level caveats and benchmark comparisons.
- vendor boosterism Google's own single-model and physical AGI framing set the pitch.
- sceptical deflation Detailed prompts, preset actions and missing failure counts undercut the demos.