Google launches Gemini 4 Argon for robotics: 87% on ERQA, reproduces papers with one prompt
keerthanpg · x · 2026-10-14
Google's Developers Blog announced Gemini 4 Argon for Robotics, a general-purpose physical reasoning model the company calls a step change toward physical AGI.
- Capability: Scores 87% on the Embodied Reasoning (ERQA) benchmark using pure reasoning, outperforming all other frontier models, with major gains in physical understanding, code generation, and visual-semantic reasoning.
- Four key application areas with representative prompts:
- Robot control: mobile manipulation with MuJoCo practice, zero-shot control, closed-loop steering of video world models;
- Real2Sim: reconstructing physically calibrated simulation twins from robot camera feeds, egocentric human video, or hand-drawn sketches;
- Physical understanding: agentic multimodal hindsight labeling, dexterity understanding, dense value functions and progress estimation;
- Research & self-improvement: action improvement via simulation and iterative self-improvement on hardware.
- Demo: controls a Boston Dynamics Spot (12-DoF legs + 6-DoF arm) to step on a foot-pedal trash bin and dispose of soda cans—first practicing in a MuJoCo digital twin, then transferring to real hardware (Real-to-Sim-to-Real).
Related event: Google Launches Gemini 4 Argon for Robotics, Nears Physical AGI(4 posts)→
More from Embodied
- Dyna-2 Robots Predict the Future: Million-Hour World Models and the Bitter Lesson of Robotics — JasonMa2020 · 2026-10-14
- AI2's MolmoAct 2 Tops Human-Preference Robot Rankings; 55K-Vote Dataset Open-Sourced — allen_ai · 2026-10-14
- Google's Gemini 4 Argon for robotics: simulation training lifts task success from 56% to 75% — DynamicWebPaige · 2026-10-14
- Founding engineer leaves YC robotics startup to build his own robot from scratch — natesiggard · 2026-10-14
- "GPT-6 Astra" demo claims 25 sim environments, 25 tasks, 5 robot embodiments from 3 prompts in 24h — abhishekunique7 · 2026-10-14
- Dyna-2: million-hour world-action model lets robots learn tasks from 45 minutes of demos — Latent Space · 2026-10-14