Gemini 4 Argon: Google DeepMind Hits 87% on Robotics Benchmark
Google DeepMind unveils Gemini 4 Argon for Robotics, scoring 87% on the ERQA benchmark and advancing Sim-to-Real learning and zero-shot physical control.
👁 112 reads
Google DeepMind has unveiled Gemini 4 Argon for Robotics - a specialized AI model designed for robotics applications that achieved an 87% score on the Embodied Reasoning (ERQA) benchmark, outpacing every other frontier model evaluated in spatial and physical reasoning. The announcement was made on October 13, 2026, by Kanishka Rao, Principal Software Engineer at DeepMind.
What is ERQA and why 87% matters
The ERQA - Embodied Reasoning and Question Answering - benchmark evaluates a model’s ability to understand and solve physical and spatial challenges in three-dimensional environments. This goes far beyond basic image recognition, focusing on complex physical reasoning: how an object behaves when dropped, where its center of gravity lies, or how a robot must position itself to avoid tipping over during intricate motions. Argon hits 87% relying strictly on pure reasoning, without relying on extra external modules or dedicated physics engines.
By comparison, earlier foundation models - even frontier systems like GPT-4o or previous Gemini iterations - struggled noticeably on such tasks. While DeepMind did not publish exact head-to-head metrics for competing models on this specific benchmark, its claim of outperforming every other frontier model is clear and verifiable. In practical robotics, this score is significant: a robot powered by such a model can evaluate grasp safety on the fly, detect trajectory collisions, and dynamically reorganize its plan when disruptions occur.
Four key application areas - from Spot to hand-drawn sketches
Argon is not narrowly limited to a single task. DeepMind highlights four core application areas across the entire robotics lifecycle.
The first is robot control. Its most striking showcase involves Argon controlling a Boston Dynamics Spot quadruped equipped with 12-degree-of-freedom legs and a 6-degree-of-freedom manipulator arm. Spot’s task was learning a brand-new skill: stepping on the foot pedal of a trash can and disposing of an empty Diet Coke can. Before executing this on physical hardware, Argon generates a digital twin of the can in the MuJoCo simulator, practicing maneuvers such as three-legged balancing while depressing the pedal. Only after this simulated training does Spot execute the task in the physical world. It uses its front camera to ensure its front-right foot is centered on the pedal, while a dual-agent setup (a foot-verifier and a hand-verifier) confirms the arm’s position over the open bin. This represents a complete Real-to-Sim-to-Real loop executed end-to-end by Argon.
The second area is Real2Sim - the automated reconstruction of physically calibrated digital twins. Argon can generate simulation environments directly from an onboard robot camera feed, egocentric human video footage, or even a simple hand-drawn sketch on paper. This drastically lowers barriers for research labs lacking the time or resources to build custom simulation environments manually.
The third area is physical reasoning, where Argon serves as an agentic multimodal hindsight labeling system. The model can judge the dexterity of an executed motion, assign dense value functions across trajectory steps, and estimate overall task progression. In practice, this automates the generation of high-quality training datasets without requiring human annotators.
The fourth area - research and self-improvement - establishes an iterative cycle where Argon suggests action refinements based on simulation outcomes, validates them, and transfers the improved control policy to physical hardware. The model thus operates not merely as an execution engine, but as an autonomous researcher.
Zero-shot control and its impact on robotics software development
Another highlight is DeepMind’s demonstration of zero-shot control. Without prior training on the target robot hardware, Argon directed a basic manipulator arm to upright an overturned cup and insert a crumpled napkin inside. The model reasoned directly from a wrist-mounted camera feed overlaid with a coordinate grid (calibrated to 2 cm per pixel in the physical world) and dispatched commands as curl requests to the robot’s API. Every few motions, it recorded a single-sentence observation, planned the next step, and executed it - bringing the architecture of digital language agents directly into physical environments.
For robotics software developers, this marks a substantial paradigm shift. Integrating large models into robotics control stacks has typically demanded complex layers of manual abstraction. Argon shows that a model can directly interpret environmental states and generate structured, machine-executable actions without requiring custom intermediate translation layers.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Physical AGI - an ambitious vision backed by practical results
DeepMind explicitly describes this milestone as accelerating progress toward “Physical AGI.” While that term is often used as marketing hyperbole, here it is grounded in tangible engineering: a model scoring 87% on an embodied reasoning benchmark, operating on actual hardware, mastering tasks via autonomous simulation, and closing the Sim-to-Real loop.
Argon is not a downloadable consumer tool today. It is a research and engineering preview showing where Google’s Gemini platform is heading in robotics. For engineering teams and robotics labs working on mobile manipulators, warehouse automation, or robotic surgery, the takeaway is clear: spatial reasoning in multimodal models has reached the threshold of practical, autonomous utility in the physical world.
And this is only the first public preview of what Argon can do.
More on this from QUANTSLANT
ai tools·Which Decision Model to Choose: A Practical GuideAnalyze top decision models by classification accuracy, tool routing, epistemic awareness, open weights, and data residency.