Gemini 4 Argon for robots: demos across the whole robotics stack

DeepMind published a robotics post about Gemini 4 Argon on October 13, positioning the model as a general-purpose physical reasoning model that can be applied across the whole robotics lifecycle. The headline number: on the ERQA embodied reasoning benchmark, which tests physical and spatial problem solving in 3D environments, Argon scores 87% using pure reasoning. According to Google, that puts it ahead of every other frontier model. The post itself is a gallery of prompts showing the model working in four areas: control, simulation reconstruction, physical understanding, and self-improvement.

Control: practice in MuJoCo, execute on hardware

The first demo puts a Boston Dynamics Spot (12 degrees of freedom in the legs, 6 in the arm) in front of a foot-pedal trash bin. Argon builds an articulated MuJoCo twin of the bin, practices stepping on it, and only then presses the real pedal to drop a soda can inside. The prompt asks for a two-agent verification swarm: one agent checks the foot step, the other the hand position over the open bin.

The zero-shot arm control is the part closest to everyday developer work. Argon drives a simple arm through a wrist camera with 2 cm world coordinates projected onto the image, writing its own action calls, curl commands against a local server, to stand up a fallen mug and drop a napkin inside it.

The third control mode steers a video world model in a closed loop. Argon writes timestamped action prompts for Gemini Omni, inspects intermediate keyframes, refines the prompts, and in a 33-step rollout rearranges five scattered Scrabble tiles into the word A-R-G-O-N.

Real2Sim: digital twins from robot footage, human video, and paper sketches

The Real2Sim section covers three inputs. From a recorded teleoperation episode, Argon reconstructs a calibrated MuJoCo scene and tunes it until the rendered cameras and the fingertip contact forces match the real data. From an egocentric video of a person working, it extracts the relevant objects, builds convex-decomposed meshes for concave containers, and retargets the hand trajectory onto a 76-degree-of-freedom humanoid with 22-degree-of-freedom hands, at 500 Hz physics. From a photo of a hand-drawn sketch: a GPU-accelerated scene where the Apollo 2 humanoid throws a ball on a parabolic arc and knocks down a 3-2-1 stack of boxes, with Google reporting 6/6 boxes down, verified in closed-loop physics.

Physical understanding: labeling data and scoring progress

On the data side, Argon annotates a 136-second overhead video of an ALOHA 2 robot folding an origami fox into a four-level hierarchy: the episode goal, 6 semantic subtasks, 17 contact-level primitives, and quality-control reweighting that downweights fumbles to 0.15x to 0.25x for behavior cloning.

Two more demos. Argon builds the classic 12-ball balance-scale puzzle as a MuJoCo environment with a two-pan Roberval scale, then solves it vision-only in three weighings. And from a 60-second bimanual UMI episode it produces a dense task-progress curve decomposed into 28 segments with positive, neutral, and negative slopes.

Self-improvement: simulation data and hardware iteration

The post's most concrete numbers come here. Starting from a single RGB photo of a real tabletop, Argon builds a calibrated simulation twin, defines a 77-task curriculum, and synthesizes more than 4,500 domain-randomized episodes. Fine-tuning a VLA model on that data lifted real-robot success rate from 56% to 75% and cut completion time from 35 s to 24 s, according to Google.

On hardware, a 6-DoF SO-101 arm improves its own pick-and-place across three iterations: 58.9 s the first, 22.9 s the second (grazing the box rim), 17.4 s the third with a centered drop. The model diagnoses its own failures from logs and camera footage. Google also shows paper reproduction: landmark locomotion and control papers ported to MuJoCo WebAssembly and running live in the browser, 68 interactive sandboxes in total, verified with headless Chromium.

Where you can call it, and what to verify

Argon is available today in Google AI Studio for AI Ultra subscribers, on the paid Gemini API tier (model gemini-4-argon, via the Interactions API), and in Google Antigravity.

Everything above comes from DeepMind's own post, and several demos ship as video. The 87% ERQA number is Google-reported and has not been independently reproduced as of the post's date. If your team works with robot simulation pipelines, the Real2Sim section is the part worth looking at. Re-run these numbers on your own workloads before putting them in any ranking.