Gemini 4 Argon for Robotics

What it claims

Gemini 4 Argon is offered as a single model for several robotics jobs: perception, code generation, simulation building, and video annotation. The headline figure is 87% on ERQA "using pure reasoning", which DeepMind says beats every other frontier model. The page lists no competitor scores. Most of the page is reusable prompt templates, not results. Those templates are the main reason to open it.

How it works

The page shows prompts and what the model is said to do with them. It does not describe the architecture or training. The recurring pattern is to give Argon a tool interface, a verification objective, and explicit rules against cheating, then let it iterate.

• Code-as-policies arm control. The robot is driven over a local HTTP API:
• GET http://127.0.0.1:8081/obs returns ee_pos, holding, status, workspace and wrist_cam_path.
• POST /plan accepts primitives: look_at, deproject(u,v,z_plane), grasp(x,y,z,yaw_deg,pitch_deg), place, place_until_contact(max_lag, relief), goto(arc_z, pitch/azimuth/roll, speed), move_by, move_until_contact, gripper open/close, home, done.
• The wrist image carries a calibrated overlay: a 2 cm grid with world coordinates in centimetres, a crosshair at the pinch centre, jaw brackets, and an object-axis line.
• On Turn 1 the model picks how fine-grained its control will be.

flowchart TD
  A[GET /obs + annotated wrist_cam] --> B{Turn 1: pick granularity}
  B -->|Coarse 2-4 calls| C[look_at → grasp → place]
  B -->|Hover verify 4-7| D[goto z=0.04, check brackets]
  B -->|Visual servoing 8-20| E[move_by / move_until_contact]
  C --> F[POST /plan, re-inspect image]
  D --> F
  E --> F
  F -->|grasp_failed| G[look_at re-observe] --> A
  F -->|ok| H[done]

• Real-to-sim twin from teleop data.
• Argon tunes object geometry, mass, friction and initial poses until two things match the real episode: the camera renders, and the fingertip contact force/torque (timing, magnitude and direction).
• Robot commands may deviate from the logged trajectory only near contact: within 0.55 rad for arm joints and 0.65 rad for hand joints.
• Explicitly forbidden: changing the model after t=0, external forces, welds, setting object qpos/qvel, disabling collisions, and reporting forces that don't come from the MuJoCo solver.
• This list is a useful template for anyone checking sim reconstructions.
• Hindsight annotation (ALOHA 2 origami, 136 s).
• Output has four levels: goal; 6 subtasks; 17 atomic primitives of 3–15 s each; and QC.
• The QC level holds task progress flat during retries and downweights them to 0.15–0.25× for behaviour cloning.
• Value function (UMI ring-sorting, 60 s, 3 camera views). Argon splits the episode into 28 segments with positive, neutral or negative progress slopes.
• Examples not in the summary:
• A bimanual Franka FR3 setup that solves the 12-ball balance-scale puzzle. The solver uses only stereo RGB (add_privileged_task_observations=False) and reads the scale tilt from vertical pixel disparity.
• A 6-DoF arm that improves itself over three stages: dry run, hover verification, then autonomous retries. Speed-ups are capped at 2× unless authorised, with safety rules in AGENTS.md.

Evidence

• ERQA: 87%, pure reasoning. No scores are given for the other models.
• VLA sim data:
• Starting from one tabletop photo, Argon reconstructed 9 props, defined a 77-task curriculum, and generated 4,500+ domain-randomised episodes with ±10 cm pose randomisation.
• Real-robot success rose from 56% to 75%, and completion time fell from 35 s to 24 s.
• The page does not name the VLA, the number of trials, or the evaluation tasks.
• Sketch-to-sim throw:
• Requirements: 6/6 boxes knocked down, pelvis drift under 3 cm, and a 76-DoF humanoid (32-DoF body plus two 22-DoF hands) in mujoco_warp.
• These are stated as requirements in the prompt. The page does not report measured outcomes.
• Paper reproduction:
• Tested on DeepLoco, HUSKY and Whole-Body MPC, ported to MuJoCo WebAssembly.
• The check is 0 console errors and non-zero task progress over 300+ frames in headless Chromium via Playwright.
• That is a smoke test, not a fidelity test.

How to use it

• The model ID is gemini-4-argon, available in the Interactions API with day-one support in the google-genai SDKs.
• The page text is cut off after that, mentioning "/goal runs, and multi-agent orchestration". No pricing, rate limits or terms are given.

Limits

• Almost everything shown is a prompt. Apart from ERQA and the VLA numbers, the page reports no success rates, trial counts or failure cases for the demos.
• ERQA has no baseline figures, and "pure reasoning" is not defined.
• The VLA improvement is a single result with no model, sample size or variance.
• The Spot, arm and self-improvement demos depend on your own harnesses: the HTTP primitive server, the calibrated overlays, and the skills such as spot-mobile-manipulation. Those harnesses are not released on this page.