RoboticsMachines in the physical world
Google launches Gemini 4 Argon for robotics; it ranks third of five on Artificial Analysis's real-arm benchmark
Google launched Gemini 4 Argon as one general model for the whole robotics workflow, from building simulations to labelling data and controlling arms. An independent real-arm benchmark ranked it third of five models, and slowest, at about 52 minutes per task.
From the Google DeepMind article of October 13, 2026 and cited sources: Argon is not a dedicated robot-action model. [3] [1]
SOURCES[3] Gemini 4 Argon for Robotics – Google Developers Blog developers.googleblog.com[1] Gemini 4 Argon for Robotics deepmind.google
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Google DeepMind article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 13, 2026 article published
- Covered
- Oct 16 edition
- Evidence
- 13 sources across 10 sites
- 1Gemini 4 Argon for Roboticsdeepmind.google
- 2Welcome to the era of embodied AI: Introducing AA-Robotics…artificialanalysis.ai
- 3Gemini 4 Argon for Robotics – Google Developers Blogdevelopers.googleblog.com
- 4Gemini Robotics: Bringing AI into the Physical World – arXivarxiv.org
- 5Introducing Gemini Robotics ER 2 – Googleblog.google
- 6Gemini 4 Argon: Built to solve your hardest problems –…blog.google
Reach: not measured yet
Argon is not a dedicated robot-action model. It works through code and tools: it builds MuJoCo scenes from camera footage, human video or sketches, labels demonstrations, estimates task progress and drives arms through a fixed set of preset motion commands. In one demo, about 4,500 simulated episodes made from a single tabletop photo were used to fine-tune a vision-language-action model, and real-robot success rose from 56% to 75%. [3] [1]
The two headline results point in different directions. Argon scored 87% on ERQA, a 400-question embodied-reasoning test that Google itself built, ahead of Claude Opus 5.5 and GPT-6 Astra. On AA-Robotics v0.1, where models directly control a real arm, it scored 55.3 out of 100 with a 36% grasp success rate, behind GPT-6 Astra's 69.4. [7] [4] [2]
Live control is still slow and costly. Argon averaged about 52 minutes per task and took roughly 21 minutes before its first action. Across the models tested, costs ran from $1.93 to $16.36 per task, and speeds were about 100 times slower than human baselines. [2]
Argon leads a question-answering benchmark but trails rivals when it actually controls an arm, where every model was slow and costly.
Read the full assessment
Implication: near-term value lies in offline work such as simulation twins, synthetic data and labelling, and teams should test models on their own hardware.
I've finished the research and am now writing the dossier.
Executive brief
Google says Gemini 4 Argon scores 87% on its own Embodied Reasoning (ERQA) benchmark and beats every other frontier model. An independent test published a day later found something different. On Artificial Analysis's AA‑Robotics v0.1, where models directly control a real robot arm, Argon placed third with 55.3 points. GPT‑6 Astra scored 69.4 (Artificial Analysis). Argon was also the slowest model, at about 52 minutes per task. Google presents Argon as a general physical-reasoning model for the whole robotics workflow: robot control, turning real scenes into simulations, labelling robot data and improving itself on hardware (Google Developers Blog). The robot demos are strong, but they depend heavily on detailed prompts and custom tooling.
What changed and event timeline
ERQA benchmark released
Google DeepMind introduced ERQA with the Gemini Robotics paper. It has 400 multiple-choice visual questions covering spatial, trajectory, action, pointing and multi-view reasoning ().
Gemini Robotics ER 2
Google released a dedicated robotics "brain" model with continuous video monitoring, available in the Gemini API ().
Argon launches with a robotics showcase
Google announced Gemini 4 Argon across the API, Antigravity and the Gemini app. A companion post showed demos on Spot, ALOHA 2, Franka and SO‑101 robots (;).
Independent real-arm benchmark
Artificial Analysis released AA‑Robotics v0.1. Argon ranked third of five models and was the slowest ().
Press analysis questions the scaffolding
The Decoder pointed to the heavily engineered prompts behind the demos and noted that Google did not report failure rates or how many attempts were made ().
Capabilities and access
Read the full section
- Model ID:
gemini-4-argon. It is available on the paid tier of the Gemini API, in AI Studio for Ultra subscribers, and in Antigravity for Ultra subscribers (Developers Blog). - Pricing: introductory rates are $2 per million input tokens and $10 per million output tokens. Standard pricing starts November 30, 2026 (Google).
- Context window: 1M tokens, according to coverage (FoneArena).
- Demonstrated skills:
- Zero-shot arm control through calls to an HTTP interface of preset motion commands (grasp, place, move until contact)
- Building MuJoCo simulation scenes from robot camera footage, first-person human video or hand-drawn sketches
- Hindsight labelling of robot demonstrations
- Estimating task progress from video (value functions)
- Reproducing robotics papers in browser-based MuJoCo WebAssembly
Technical analysis for researchers and developers
- No dedicated robot-action model. Argon is a general model that acts through code and tools.
- Prompts carry much of the engineering. Real-to-sim fidelity is checked against recorded fingertip forces.
- Prompts and outputs are published, but weights and evaluation code are not. Trial counts behind the success rates are also missing (The Decoder).
Read the full section
- No dedicated robot-action model. Argon is a general model that acts through code and tools. It writes MuJoCo scenes, inverse-kinematics trajectory generators and solvers, then checks them in physics simulation. In the zero-shot demo it drives the arm with a calibrated wrist-camera grid and a fixed set of motion commands (Developers Blog).
- Prompts carry much of the engineering. The real-to-sim prompts include explicit "no sim cheating" rules: no welds, no applied external forces, no copying logged tactile data. Real-to-sim fidelity is checked against recorded fingertip forces.
- Prompts and outputs are published, but weights and evaluation code are not. Trial counts behind the success rates are also missing (The Decoder).
- Method details from the independent test: AA‑Robotics ran Argon through Antigravity at high effort, on a Hiwonder 6-axis arm with four cameras. Partial task progress earned partial credit (Artificial Analysis).
Claims and evidence
- ERQA: Argon 87%, Claude Opus 5.5 82.3%, GPT‑6 Astra 81%, run with no tools (The Decoder).
- Simulation-generated training data: about 4,500 episodes generated from a single tabletop photo were used to fine-tune a vision-language-action (VLA) model.
- Cornhole demo: Argon watched a human cornhole video and built practice simulation code.
Read the full section
- ERQA: Argon 87%, Claude Opus 5.5 82.3%, GPT‑6 Astra 81%, run with no tools (The Decoder). Google built both the benchmark and the model, and no independent rerun has been found.
- Simulation-generated training data: about 4,500 episodes generated from a single tabletop photo were used to fine-tune a vision-language-action (VLA) model. Real-robot success rose from 56% to 75% and completion time fell from 35s to 24s (Developers Blog). This is a vendor result with unreported trial counts.
- Cornhole demo: Argon watched a human cornhole video and built practice simulation code. Transferred to Apptronik's Apollo humanoid, the robot made seven shots in a row (Google on X).
- Real-arm control: independently scored at 55.3/100 with a 36% grasp success rate (Artificial Analysis).
Context and prior work
ERQA dates from the March 2025 Gemini Robotics work (arXiv). Gemini Robotics 1.5 added a thinking-based embodied-reasoning model (arXiv), and ER 2 followed in July 2026 (Google blog). Argon's demo setup, with preset motion commands called through HTTP, follows the earlier "code-as-policies" approach of having models write robot control code.
Read the full section
ERQA dates from the March 2025 Gemini Robotics work (arXiv). Gemini Robotics 1.5 added a thinking-based embodied-reasoning model (arXiv), and ER 2 followed in July 2026 (Google blog). Argon's demo setup, with preset motion commands called through HTTP, follows the earlier "code-as-policies" approach of having models write robot control code. What Google presents as new is a single general frontier model running the whole loop, from simulation to labelling to control.
Limitations, safety and contested findings
- Argon leads Google's ERQA question-answering test but trails two GPT models when it actually controls a real arm (Artificial Analysis).
- Speed and hesitation: Argon averaged 52 minutes per task and spent about 21 minutes before its first action.
- Demo selection: Google did not disclose failure rates (The Decoder).
Read the full section
- The two headline results disagree. Argon leads Google's ERQA question-answering test but trails two GPT models when it actually controls a real arm (Artificial Analysis).
- Speed and hesitation: Argon averaged 52 minutes per task and spent about 21 minutes before its first action. Across all models, AA reports performance roughly 100x slower than human baselines and calls them far too slow for production. Models also kept refining attempts after they had already succeeded.
- Demo selection: Google did not disclose failure rates (The Decoder).
- Safety disclosures are general, not robotics-specific. Google's launch post covers bio red-teaming and prompt-injection results, but nothing about physical safety (Google).
- Community concerns: Hacker News commenters raised on-robot power limits and the risk of hacked robots (HN).
Business and practitioner implications
- The best-supported uses are building simulation twins, generating synthetic training data and labelling or quality-checking demonstrations. These run offline, so speed matters less.
- Not ready for real-time control. AA measured $1.93–$16.36 per task across models and about 30-minute tasks (Artificial Analysis).
- Results on vendor benchmarks did not match independent real-arm rankings. Teams should test models in their own setups, with failure counts logged.
Read the full section
- Most useful near term: offline work. The best-supported uses are building simulation twins, generating synthetic training data and labelling or quality-checking demonstrations. These run offline, so speed matters less.
- Not ready for real-time control. Live robot control by frontier language models is costly and slow. AA measured $1.93–$16.36 per task across models and about 30-minute tasks (Artificial Analysis).
- Test before you choose. Results on vendor benchmarks did not match independent real-arm rankings. Teams should test models in their own setups, with failure counts logged.
- Budget for prompt and harness engineering. Much of the demonstrated capability sits in the prompts, calibration overlays and motion-command libraries.
The source trail.
Sources (13)
Gemini 4 Argon for Robotics
Article text retrieved; extracted text may omit tables or interactive elements.
deepmind.google