Skip to content
Bud's Wiki

Gemini 4 Argon for Robotics

TL;DR

Section titled “TL;DR”

Google DeepMind’s first-party positioning of Gemini 4 Argon (Google DeepMind announces Gemini 4 Argon — new frontier model with 1M-token output for coding, enterprise knowledge work, and cybersecurity defense) as a general-purpose physical reasoning model across the robotics lifecycle, with 11 end-to-end prompt + task walkthroughs organized under four areas (control, world modeling for data, data generation / augmentation, self-improvement + reproducibility). The headline capability claim is 87% on the Embodied Reasoning (ERQA) benchmark via pure reasoning, outperforming “every other frontier model.” The post sidesteps the VLA-vs-WAM recipe debate entirely by positioning Argon as the orchestrator that drives everything else — writing MuJoCo scenes, picking primitives from a code-as-policies interface, steering Gemini Omni world-model rollouts, authoring digital twins from teleop episodes, retargeting human motion to 76-DoF humanoids with 22-DoF hands, running dense value functions, iteratively self-improving on hardware, and reproducing papers into interactive WebAssembly labs. Available today via gemini-4-argon in the google-genai SDKs.

Key claims

Section titled “Key claims”
  • Argon scores 87% on ERQA via pure reasoning, claimed to outperform every other frontier model on this 3D spatial-reasoning benchmark [intro §].
  • Argon drives a 12-DoF-legs + 6-DoF-arm Boston Dynamics Spot to a new skill (foot-pedal trash disposal) by authoring an articulated MuJoCo digital twin of the pedal first and training on it before the real-hardware attempt, with a 2-agent verification swarm (foot-verifier + hand-verifier) gating pedal step and bin alignment [Spot §].
  • A single prompt specifies a closed-loop code-as-policies interface for a 6-DoF arm (robot.look_at / grasp / place_until_contact / move_until_contact / goto / move_by) with annotated wrist-cam visual coordinates and three selectable VLM-call granularities (coarse / pre-grasp-hover-verify / fine visual-servoing); Argon picks granularity upfront on turn 1 [mug §].
  • Argon steers Gemini Omni (video world model) in a closed visual feedback loop: inspects keyframes from the latest rollout and writes the next timestamped action prompt inside <omni_prompt> tags to iteratively arrange 5 Scrabble tiles into A-R-G-O-N [Omni §].
  • From a single real robot teleoperation episode Argon builds a 500 Hz MuJoCo digital twin with sim rendering that matches real head/left-wrist/right-wrist cameras and simulated fingertip force/torque that matches logged tactile data in timing, magnitude, and direction — under explicit “no sim cheating” constraints (object motion only via mj_step contact, no forces/welds/qpos edits, ±0.55 rad arm / ±0.65 rad hand command deviation from logged teleop) [digital-twin §].
  • Argon retargets an egocentric human manipulation video onto a bimanual humanoid with two 22-DoF five-fingered hands at 500 Hz with convex-decomposed meshes for concave containers, verifying thumb-finger force closure, >2 cm lift, and stable placement — no kinematic weld constraints [egocentric §].
  • Argon translates a pen-and-paper sketch of a 76-DoF humanoid throwing at a 3-2-1 box stack into a verified mujoco_warp scene + 50 Hz four-phase throw trajectory (grasp, wind-up, whipping, synchronized 22-DoF finger release), verifies 6/6 knockdown with <3 cm pelvis drift, and exports a Blender 4.5 Cycles render [sketch §].
  • On an ALOHA 2 origami fox folding video (136 s), Argon produces a four-level time-aligned hindsight annotation: G1 (episode goal), G2 (6 contiguous semantic subtasks with which arm anchors vs folds/rotates), G3 (17 contact-level atomic primitives aligned with grasp closure / crease sweeps / release / retry attempts), plus QC reweighting that holds task_progress_pct flat during slip/delamination retries and downweights them 0.15×–0.25× for behavior cloning [ALOHA §].
  • Argon implements a 12-ball balance-puzzle MuJoCo env from scratch (12 visually-identical pool balls with one 270 g or 30 g odd ball, articulated Roberval two-pan scale with mjEQ_JOINT parallelogram constraints, Red/Blue Heavy/Light submission trays) and writes a vision-only closed-loop solver that triangulates balls from stereo RGB, visually reads scale tilt from pixel disparity, rotates wrist yaw 90° to clear pan rims, and isolates the odd ball in ≤3 weighings [balance §].
  • Argon acts as a general value function for robotics tasks: on a 60 s bimanual ring-sorting UMI human-video episode (3 synchronized views) it iteratively inspects contact transitions and decomposes the trajectory into 28 segments with positive, neutral, and negative progress slopes [value-function §].
  • VLA sim-data augmentation: starting from a single RGB photo of a real tabletop, Argon autonomously builds a calibrated MuJoCo twin (procedural meshes + compound collision primitives, spring-hinged deformable crown panels for a plush cap), defines a 77-task single-arm + bimanual curriculum, synthesizes a closed-loop IK trajectory generator across ±10 cm pose randomization, verifies reward == 1.0, and exports 4,500+ episodes that lift real-world Franka success 56% → 75% and reduce completion time 35 s → 24 s [VLA-sim §].
  • Iterative self-improvement on a real 6-DoF arm: three-stage prompt (measure + plan dry-run → pre-pick hover + wrist-cam verify → execute + self-diagnose from own logs + make ≥2× faster subject to safety rules) runs autonomously with Argon reviewing its own wrist/overhead footage to find root causes and retry [self-improvement §].
  • Paper reproduction into browser labs: Argon reads a robotics-control manuscript + released checkpoints and ports the physics model + closed-loop controller (named examples: DeepLoco’s 7-network hierarchical locomotion, HUSKY’s 50 Hz PPO actors, Whole-Body MPC’s Web Worker iLQR) into self-contained MuJoCo WebAssembly browser labs, verified in headless Chromium with 0 console errors over 300+ frames [paper-repro §].

Method

Section titled “Method”

The post is not a tech report — no architecture, parameter count, training compute, model-card link, or quantitative head-to-head beyond the ERQA 87% number is disclosed. What it does ship is 11 full prompt specifications (reproduced verbatim in the post body) that document a consistent interaction pattern: Argon receives a goal plus multimodal context (RGB images, egocentric video, hand-drawn sketches, URDF-level robot specs, point-cloud or scene files, teleop logs with tactile/force channels), then emits either (a) tool calls into a provided code-as-policies / robot-control API, (b) typed JSON annotations, (c) MuJoCo / mujoco_warp / Blender XML and Python, or (d) prompts for a downstream video world model (Gemini Omni). Every prompt specifies a verification objective and most specify an iterative refinement loop — Argon self-checks keyframes, cross-camera agreement, contact wrenches, task progress, or physics-engine ground-truth before advancing. The 500 Hz dexterous retargeting and mujoco_warp humanoid throw demos are the only ones that explicitly cross the sensorimotor-contact threshold Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level where previous Claude- and Gemini-driven tool-use agents have struggled on real hardware (Gemini Robotics-ER 2 vs Claude Opus 5 clapboard eval — ER2 threw robot arms off table, hallucinated success (Robocurve / Jay Chooi), Comparing Fable 5.1 to Fable 5 on robotic manipulation).

Results

Section titled “Results”
  • ERQA: 87% via pure reasoning (benchmark leader claim) [intro].
  • VLA sim-data pipeline: real-world Franka success 56% → 75%, completion time 35 s → 24 s on an unspecified 77-task curriculum from a 4,500+-episode sim-generated dataset [VLA-sim §].
  • Spot foot-pedal trash task: qualitative demonstration, no success rate reported [Spot §].
  • Humanoid sketch-to-throw: 6/6 target knockdown with <3 cm pelvis drift in mujoco_warp physics, qualitative Blender Cycles render exported [sketch §].
  • All other sections (code-as-policies mug, Gemini Omni Scrabble, digital-twin from teleop, egocentric retargeting, ALOHA origami annotation, balance puzzle, general value function, iterative self-improvement, paper reproduction) are qualitative demonstrations without reported success rates, latencies, or baselines against other models.

Why it’s interesting

Section titled “Why it’s interesting”

This post is the most explicit first-party statement to date that a frontier general-purpose LLM — not a specialized VLA — is now positioned as the orchestrator-of-everything across the robotics lifecycle. It sits directly on the “LLM-as-controller” row of the VLA Models recipe-lever board that papers like OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies OpenRUA and Agent as Policy for Robotic Manipulation AGP have been mapping empirically, with Argon now representing Google’s bet on the extreme end of that axis (ROS 2 / native-tool-calls / no VLA in the inner loop). The sim-data augmentation result (56→75% VLA real-world success) is a load-bearing new datapoint for Synthetic Training Data and the Simulator-in-the-Loop Refinement cell. It also strengthens the World Foundation Models “WFM-as-companion-steered-by-LLM” pattern — Argon as the Omni-steering outer loop looks closest to SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators SyncWorld’s training-free rank-and-execute loop but with iterative prompt rewriting as the lever rather than best-of-N sampling. The ERQA 87% headline, if it holds against the harder spatial-reasoning splits other filed papers measure (Cambrian-S: Towards Spatial Supersensing in Video, Solving Spatial Supersensing Without Spatial Supersensing), would be the strongest filed challenge yet to the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position’s prediction that language-model scaling can’t close the gap on contact-rich 3D reasoning.

Companion piece to Google’s Gemini Robotics 2 brings whole body intelligence to robots Gemini Robotics 2 release: GR2 is the VLA, Argon is the orchestrator above it — the two-tier recipe GR2’s blog named as “ER + VLA” now explicitly instantiated by swapping GR2-ER for a general frontier model. Zero disclosure of architecture or training compute keeps it in the same transparency regime as the Argon announcement — flagged for /bud refresh once a model card or technical report lands.

See also

Section titled “See also”