Google DeepMind unveils Gemini 4 Argon, a physical reasoning model for robotics

Google DeepMind has published a blog post presenting Gemini 4 Argon, which it calls a general-purpose physical reasoning model that can be applied across the entire robotics lifecycle. The post claims Argon marks a step change in physical understanding, code generation and visual-semantic reasoning, unlocking new workflows across the robotics stack and accelerating progress toward physical AGI.

The headline number is on the Embodied Reasoning (ERQA) benchmark, which the post says tests a model's ability to understand and solve physical, spatial problems in 3D environments. Using pure reasoning, Argon scores 87%, which DeepMind says outperforms every other frontier model. The post says it shares representative prompts across four key areas; the visible text is cut off mid-article, so the rest of the post is not covered here.

The demos are shown as prompts plus short descriptions. In the first, Argon controls a Boston Dynamics Spot quadruped (12-DoF legs plus a 6-DoF arm) to learn a new skill: stepping on a foot-pedal trash bin to throw away soda cans. Argon builds an articulated MuJoCo digital twin of the bin to practice its moves before stepping on the real pedal. The prompt asks Spot to pick up two Diet Coke cans and bin them, to check in the front cameras that its front-right foot is centered on the pedal, and to use a 2-agent verification swarm (a foot-verifier and a hand-verifier) before releasing.

In the second, Argon controls a simple robotic arm to stand a fallen mug upright and put a crumpled napkin inside. It works over several turns with a visual coordinate harness (a calibrated grid and reticle drawn on the wrist camera image) and action primitives such as look_at, grasp, place, goto, move_by and move_until_contact, sent as curl commands to a local /plan endpoint. The prompt lets the model choose its call granularity on turn 1: coarse (2 to 4 calls), pre-grasp hover verify (4 to 7 calls) or fine visual servoing (8 to 20 calls).

In the third, Argon steers a video world model, the Gemini Omni World Model, to rearrange five scattered Scrabble tiles (G, R, A, N, O) into a row spelling A-R-G-O-N. Rather than one static prompt, it works in a closed visual loop: it writes timestamped action prompts in [MM:SS.mmm] format, inspects intermediate rollout keyframes and refines the next prompt, breaking motion into single-direction actions.

In the fourth, Argon watches a real robot teleoperation episode and builds a digital twin. It creates or obtains simulation assets that look and behave like the real objects on contact with dexterous hands, then tunes assets and scenes in an iterative visual-reasoning loop. The prompt sets two objectives: a visual match across the head_left, left_wrist and right_wrist cameras, and a physical match of simulated fingertip contact force and torque against the logged tactile_force in timing, magnitude and direction. Tunable items are static object assets (geometry, collision shapes, mass, inertia, friction, joints, initial poses) and offline-planned robot commands, changed only where fingertips are near an object and kept within 0.55 rad for arm joints and 0.65 rad for hand joints of the logged trajectory. The prompt bans cheating: no external forces, welds, qpos/qvel setting, disabled collisions, or copying the logged tactile readings. The result is to be shown in an interactive HTML session viewer with a synced timeline, a 3D Blender scene and real versus sim camera and force views.

In the fifth, given an egocentric human video of a manipulation task, Argon identifies the task-relevant objects, replaces crude bounding boxes with compound convex-decomposed meshes for concave containers, and retargets the human hand trajectory onto a 76-DoF bimanual humanoid with 22-DoF five-fingered hands in MuJoCo contact physics. The prompt asks for a 500 Hz environment (dt = 0.002 s) with no kinematic weld constraints, and checks for thumb-and-finger force closure, a clean lift of more than 2 cm and stable final placement.

In the sixth, a developer photographs a rough pen-and-paper sketch of an Apollo 2 humanoid throwing a ball along a parabolic arc to topple a 3-2-1 stack of boxes. Argon then builds the GPU-accelerated mujoco_warp scene, synthesizes a balanced 76-DoF whole-body throw (a 32-DoF body plus two 22-DoF hands) as a 50 Hz trajectory of 140 steps, or 2.80 s, across four phases, verifies 6/6 target knockdown in closed-loop physics with under 3 cm of pelvis drift, and exports a Blender 4.5 Cycles render.

The last demo visible is video annotation. Given an overhead video of a bimanual ALOHA 2 robot folding a square sheet of paper into an origami fox over 136 seconds, Argon produces a time-aligned four-level hindsight annotation hierarchy: an episode goal (G1), 6 contiguous semantic subtasks (G2), 17 contact-level atomic primitives (G3), and automated quality-control reweighting that holds task progress flat during fumbles and retries. The post says Argon can identify poor manipulation segments such as pauses, unstable grasps and fumbles. The prompt for this demo is cut off in the visible text.

Key facts

• DeepMind says Gemini 4 Argon scores 87% on the Embodied Reasoning (ERQA) benchmark using pure reasoning, outperforming every other frontier model.
• Argon is described as a general-purpose physical reasoning model for the whole robotics lifecycle, shown through example prompts rather than results tables.
• Demos include a Boston Dynamics Spot (12-DoF legs plus 6-DoF arm) learning to step on a foot-pedal bin after practising on a MuJoCo digital twin.
• A sketch-to-simulation demo builds a mujoco_warp scene for a 76-DoF Apollo 2 humanoid and verifies 6/6 box knockdown with under 3 cm pelvis drift.
• Argon also annotates a 136-second ALOHA 2 origami video into 1 goal, 6 subtasks and 17 atomic primitives, flagging pauses, unstable grasps and fumbles.

Why it matters

This is a first-party launch of a model aimed squarely at robotics, and the pitch is breadth: one model said to cover controlling robots, building simulations, retargeting human video and labelling robot data. DeepMind frames it as a step change in physical understanding, code generation and visual-semantic reasoning and as progress toward physical AGI. Much of the work is done by writing code and scene descriptions (MuJoCo, mujoco_warp, Blender) that are then checked in physics simulation, and by steering a video world model in a loop, rather than by a single model emitting motor commands.

Who it affects

Robotics developers and researchers are the target audience. The demos touch teams working with quadrupeds like Spot, simple robotic arms, bimanual platforms like ALOHA 2, and humanoids with dexterous five-fingered hands. People who build simulation environments, digital twins and robot datasets are addressed directly, since several demos replace manual scene building and manual video annotation with prompts.

How to use it

The post is mainly a showcase of representative prompts. They show patterns: attach images, video and JSON observations; specify tight verification criteria (tactile force match, more than 2 cm lift, under 3 cm pelvis drift, 6/6 knockdown); state explicit rules against simulator cheating; and let the model work in closed loops, inspecting renders or camera frames after each step. No release date, availability, pricing or API access for Gemini 4 Argon is given in the visible text.

How solid is it

The 87% ERQA figure is DeepMind's own claim, in a first-party post. The other frontier models compared are not named and no competitor scores are given. The ERQA 87% figure is not stated to be a pass@1, average or best-of-n result beyond 'using pure reasoning'. No success rates, failure rates or number of trials are given for any of the demos; the text shows prompts and descriptions only. The source text supplied is truncated mid-article, so later sections are not seen.

Risks and caveats

Treat the demos as illustrations, not measured results: with no trial counts, it is unclear how reliable any of them is. The visible text does not say whether the Spot, robot-arm and teleoperation demos were run end to end on real hardware versus in simulation beyond what the descriptions state. No model size, architecture, training data or parameter count is stated. The post says the model 'accelerates progress toward physical AGI', which is the authors' own framing, not an independent finding. The visible text names no authors or quoted individuals.

“Using pure reasoning, Argon scores 87%, outperforming every other frontier model.”

— Google DeepMind blog post on Gemini 4 Argon