Gemini 4 Argon for Robotics

What it claims

Gemini 4 Argon, available as gemini-4-argon, is pitched as a general physical-reasoning model. The post is mostly a set of prompt templates showing how Google drives it in agentic loops: it emits code or prompts, observes rendered or real camera frames, verifies, and retries. The headline number is 87% on ERQA "using pure reasoning", described as ahead of every other frontier model. The post gives no competitor scores.

How it works

The common pattern in the templates is to give Argon tools, an observation channel, and explicit verification criteria, then let it iterate.

• Code-as-policies arm control. A local HTTP harness runs on 127.0.0.1:8081.
• GET /obs returns ee_pos, holding, status, workspace and wrist_cam_path.
• POST /plan takes Python-style 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 and done.
• The wrist image carries a drawn 2 cm world grid, a pinch-center crosshair, jaw brackets and an object-axis line, so the model reads coordinates visually.
• On turn 1 the model chooses a call budget: coarse (2–4 calls), hover-verify (4–7) or fine visual servoing (8–20). On grasp_failed it recovers with look_at.
• Anti-cheating constraints for real-to-sim twins.
• Object motion may come only from mj_step contact.
• After t=0 the model may not change MjModel, apply external forces, add welds or equality constraints, reset qpos/qvel, disable collisions, or copy logged tactile readings.
• Robot command edits must stay within 0.55 rad (arm) and 0.65 rad (hand) of the teleop log, and only near contact.
• Fit is judged on three-camera visual match plus Sharpa fingertip force timing, magnitude and direction.
• World-model steering. For Gemini Omni, Argon writes [MM:SS.mmm] timestamped single-direction action beats inside <omni_prompt> tags. It states explicitly which workspace areas become empty, and refines the beats after viewing rollout keyframes.
• Hindsight annotation for behavior cloning (BC). Argon labels data at four levels: goal, subtasks, 3–15 s contact primitives, and a QC pass. The QC pass holds task_progress_pct flat during retries and downweights those segments to 0.15–0.25× for BC.
• Staged real-hardware autonomy (6-DoF arm). The arm first does a no-motion dry run with an annotated plan. It then hovers above the target with wrist-camera verification. Finally it executes, self-diagnoses failures from logs and footage, and speeds up at least 2× within safety rules from an AGENTS.md.

Evidence

• ERQA: 87% with pure reasoning. No baseline numbers are given.
• Sim-data for VLA fine-tuning: Starting from a single overhead photo, Argon built a twin of 9 props and a 77-task bimanual curriculum. It collected 4,500+ episodes with ±10 cm pose randomization, keeping only reward == 1.0 episodes. Real-robot success rose from 56% to 75% and completion time fell from 35 s to 24 s. The setup was two Franka FR3 arms. The post does not name the VLA, the trial counts or the tasks tested.
• Other demos, each reported qualitatively:
• 12-ball weighing puzzle solved vision-only with stereo triangulation, ≤3 weighings.
• Sketch-to-mujoco_warp humanoid throw toppling 6/6 boxes with <3 cm pelvis drift.
• 28-segment value function on a 60 s UMI episode.
• Paper reproductions (DeepLoco, HUSKY, Whole-Body MPC) in MuJoCo WASM, checked by Playwright for 0 console errors over 300+ frames.

How to use it

The model ID is gemini-4-argon, served through the Interactions API with day-one support in the google-genai SDKs. The page text is truncated around further platform details. The prompts above are the reusable artifact. The harnesses they assume are not shown as released: the :8081 arm server, Spot "spot-mobile-manipulation" skill, verification swarms and Gemini Omni access.

Limits

• The post gives no pricing, rate limits or latency, and no failure rates for any demo.
• Apart from the VLA uplift, results are single showcase episodes rather than evaluations.
• The ERQA claim lacks comparison numbers.
• Readers should confirm whether the robot-control harnesses and Gemini Omni are actually available before planning around these workflows.