Google Shows Gemini 4 Argon Controlling Robots—but Publishes No Reliability Study

Google published an Argon robotics package on October 13

Google’s October 13, 2026 post presents Gemini 4 Argon as a general-purpose physical-reasoning model across four areas: robot control, Real2Sim reconstruction, physical understanding, and research or self-improvement. The callable product is the same gemini-4-argon endpoint covered in Google’s broader launch—not a dedicated robotics model, controller or managed robotics API.

The paid Gemini API is available now through AI Studio. Google AI Ultra subscribers can also use Argon in AI Studio and Antigravity. Introductory API pricing is $2 per million input tokens and $10 per million output tokens through November 30; the developer guide says those rates rise to $4 and $20 on December 1. Google does not publish a robot-task price because the surrounding cameras, simulation, control harness, compute and hardware remain operator-owned.

The demonstrations expose more engineering detail than a highlight reel

Google includes long representative prompts rather than only edited videos. One Spot workflow asks Argon to build a MuJoCo twin of a foot-pedal bin, practise three-leg balance, verify foot and hand alignment with two agents, then dispose of cans. A separate arm example gives the model calibrated wrist-camera coordinates and explicit look, grasp, move, place, contact and recovery primitives. Another loop has Argon steer a video world model by inspecting rollout frames and issuing timestamped action prompts.

The Real2Sim examples ask the model to reconstruct scenes from robot video, egocentric human video and a paper sketch. The prompts specify camera matching, contact-force matching, collision constraints, trajectory limits and physical success checks. These details make the post useful for architecture review: the apparent autonomy depends on instrumented observations, constrained action vocabularies, simulators, verification loops and substantial task-specific scaffolding.

The 87% ERQA score is not reproduced by the linked artifact

Google says Argon scores 87% on Embodied Reasoning Question Answer, or ERQA, and outperforms every other frontier model using pure reasoning. The linked public repository describes 400 multimodal, interleaved image-and-text multiple-choice questions about spatial reasoning and world knowledge. It provides a TFRecord dataset, loader and lightweight API harness under CC-BY-4.0.

The repository’s pinned README and eight-commit history predate Argon. They do not include an Argon model configuration, exact prompt, decoding settings, per-question outputs, comparison table or 87% result. AccessAllGPT therefore treats 87% and the ranking statement as Google-reported claims. The public dataset makes an independent run possible in principle, but reproducing the launch number requires the missing evaluation choices as well as paid model access.

The robot results need denominators and failure records

Google reports that one workflow generated more than 4,500 domain-randomized simulation episodes, improved real-world robot success from 56% to 75%, and reduced completion time from 35 to 24 seconds. It also shows a three-iteration visual pick-and-place sequence falling from 58.9 seconds to 22.9 seconds and then 17.4 seconds; the second attempt grazed the box rim. These are concrete numbers, but the page does not state the number of held-out tasks or trials, confidence intervals, selection rule, baseline training details or complete failure distribution.

Other demonstrations are presented as successful videos with prompts but no repeated-trial statistics. The post does not report intervention rate, collision or near-miss count, out-of-distribution objects, camera perturbations, model-call latency, token use, API errors, simulator-to-reality failures or emergency-stop activations. A selected completion can establish feasibility of that instrumented path; it cannot establish dependable autonomous operation.

Model reasoning must stay above a certified control boundary

Argon can propose plans, generate simulation assets, label trajectories and choose among bounded motion primitives. It should not own the final safety envelope. Workspace limits, joint and torque limits, collision checking, speed and force limits, interlocks, watchdogs, emergency stops and access control need deterministic implementations that reject unsafe model output. Verification agents generated by the same model are not independent safety channels.

Keep simulation and physical execution as separate approval stages. Require a human or qualified supervisory system to approve new assets, controllers and task classes; log every observation, model response, parsed command and controller result; and fail closed when cameras, calibration, network access or model responses are missing or stale. The right comparison is accepted task completion under fixed safety limits, not whether a polished demo eventually succeeds.

What robotics teams should do next

Start with a read-only or simulation-only evaluation using the paid gemini-4-argon API. Freeze the prompt, model identifier, thinking level, simulator, assets and controller version. Build a representative case set that includes object shifts, occlusion, glare, calibration error, failed grasps, stale observations and unreachable goals. Measure success by task, interventions, unsafe proposals, controller rejections, collisions, latency, tokens, retries and complete cost per accepted run.

Advance to hardware only behind low speed and force limits, a small workspace, independent collision and stop systems, and a named operator. Trial when the task-specific harness and safety case are inspectable. Constrain the model to planning or labeling when physical reliability is unresolved. Wait if API latency or data policy cannot meet the control loop. Reject direct execution when the model can bypass deterministic limits or when selected videos are the only evidence.

Disclosures

AccessAllGPT received no Google access, credits, hardware, data, code, outputs, benchmark logs, briefing, review, payment or compensation for this article. Google and the cited project authors did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank the model. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Google, Google DeepMind, Boston Dynamics or organizations cited. Publication-wide relationships are listed on the disclosures page.