Gemini 4 Argon Posts 87% ERQA for Robot Control

Gemini 4 Argon is live as Google DeepMind’s physical-reasoning model for robotics, and the company’s own Embodied Reasoning (ERQA) figure is 87% under pure reasoning. In the companion demo, the model maps physics from video, writes its own simulation, and teaches Apptronik’s Apollo humanoid to sink seven cornhole throws in a row.

The Spot foot-pedal trash workflow has Argon build an articulated MuJoCo twin and practice before the real pedal step. The Google Developers Blog post by Kanishka Rao carries a Oct. 13, 2026 dateline. The seven-throw clip sits on Google DeepMind’s X announcement.

Argon scores 87 percent ERQA and outperforms every other frontier model Google compares.

ERQA, as Google describes it, tests whether a model can understand and solve physical, spatial problems in 3D environments. Argon posts 87% on that mix and, in Google’s wording, outperforms every other frontier model Google compares against in that post. The score is a vendor-run leaderboard row, not a third-party ledger with shared prompts and graded transcripts.

Access already trails the capability claim. Google AI Studio and Google Antigravity expose Argon to Google AI Ultra subscribers. The paid Gemini API tier takes the model id gemini-4-argon through the Interactions API with day-one google-genai SDK support. The robotics post lists Ultra and paid Gemini API access but does not publish input or output dollar rates per million tokens.

Update (Oct 14, 2026): Google’s main Oct 13 Gemini 4 Argon launch post does publish API rates: an introductory $2 per 1M input tokens and $10 per 1M output tokens, with cached input 95% off, rising to $4 and $20 after the intro period ends Nov 30, 2026. It also puts Argon in the Gemini app and AI Mode in Search for Google AI Pro subscribers, not just Ultra. See our Gemini 4 Argon tool page for the full breakdown.

The control path builds MuJoCo twins before any real pedal or gripper commits.

Argon’s robotics workflow is a closed loop over cameras, proprioception, and simulator code rather than a single open-loop policy dump. On Boston Dynamics Spot, the model invents an articulated MuJoCo twin of a foot-pedal trash bin, practices tripod balance and pedal centering in simulation, then runs a two-agent verification swarm (foot-verifier and hand-verifier) before the real can drop. On a tabletop arm, it reasons over a calibrated wrist-camera harness with a 2 cm world grid and issues motion primitives through a local plan endpoint: look_at, grasp, place_until_contact, move_by, move_until_contact, with curl-shaped turns that re-check wrist frames after every step.

Real2Sim is the same loop pointed at reconstruction. From a robot teleop episode, Argon rebuilds assets so MuJoCo camera views and fingertip force/torque line up with logged tactile events, under hard bans on post-t=0 model edits, external wrenches, welds, and fake contact reports. From egocentric human video, it swaps crude boxes for convex-decomposed meshes and retargets hand keypoints onto a 76-DoF bimanual humanoid with 22-DoF five-finger hands at 500 Hz MuJoCo stepping. From a pen sketch of an Apollo 2 humanoid throw, it builds a mujoco_warp scene, synthesizes a 50 Hz whole-body throw (140 steps over 2.80 s), and checks upright balance under a 3 cm pelvis-drift cap while verifying a 6/6 box knockdown. Steering Gemini Omni world-model rollouts uses the same discipline: timestamped <omni_prompt> beats, one simple action at a time, inactive hands retracted so the camera still sees the workspace.

That machinery is expensive in operator attention even when token prices stay opaque. Prompts ship multi-page execution discipline, stereo triangulation rules, and explicit “no sim cheating” clauses. A lab that cannot host MuJoCo, wrist calibration, and a client loop will not get the demos Google filmed. Argon also does hindsight labeling on long ALOHA 2 origami video (G1 goal, G2 subtasks, G3 contact primitives, QC reweighting that flattens progress during fumbles) and can stand up a vision-only 12-ball balance-scale solver whose puzzle statement Google points at the classic writeup on Math is Fun.

Google’s own tabletop run lifts success from 56 percent to 75 percent under vendor measurement.

The headline hardware improvement in the blog is narrower than the ERQA brag and clearer about the method. From a single overhead RGB photo of a bimanual Franka FR3 party-prop tabletop, Argon builds a calibrated MuJoCo twin, defines a 77-task curriculum, and synthesizes a closed-loop motion generator that collects 4,500+ domain-randomized sim episodes for VLA fine-tuning. Google reports real-world success rising from 56% to 75% and completion time falling from 35s to 24s on that evaluation. Separately, an SO-101 pick-and-place self-improvement sequence shows wall-clock drops across iterations 1→2→3 (58.9s, then 22.9s with a rim graze, then 17.4s centered), again as vendor-filmed hardware, not an external bake-off.

Those percentages buy a reader proof-of-method on Google’s props, cameras, and reward definition. Google measured Google’s stack on Google’s task mix. The cornhole demo, per Google DeepMind’s X post, shows Argon teaching Apptronik’s Apollo to sink seven throws in a row after watching a video, mapping physics, and coding its own simulation.

Paper reproduction is another vendor-controlled stress test. Argon reads a robotics manuscript plus optional checkpoints, ports physics and controllers into MuJoCo WebAssembly browser labs, and verifies headless Chromium runs with zero console errors and non-zero task progress over 300+ frames. Google cites 68 interactive sandboxes spanning DeepLoco-style hierarchical locomotion, ANYmal-scale parallel RL, Whole-Body MPC, and HUSKY skateboarding. The launch still leaves industrial buyers without a shared public scoreboard they can re-run on their own robots tonight.

Outside labs still lack a public ERQA itemization they can regrade.

Ultra seats and paid Gemini API callers can select gemini-4-argon today for planning, /goal-style runs, and multi-agent orchestration inside Antigravity. Robotics shops with MuJoCo pipelines and wrist calibration are the fit. Teams waiting for open weights or on-prem guarantees do not have those in the launch posts; the per-token rates are in Google’s main launch post, linked above.

Google has not released a third-party-regradable ERQA prompt set with per-item outcomes that would let an outside lab reproduce the 87% number, and the robotics launch post itself does not publish dollar rates per million tokens for gemini-4-argon (Google’s main launch post does).