Gemini Robotics brings AI into the physical world

Source: DeepMind Date: 2025-03-12 URL: https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/

Summary

Google DeepMind launched Gemini Robotics and Gemini Robotics-ER (Extended Reasoning), both vision-language-action models built on Gemini 2.0. Gemini Robotics more than doubles performance on a generalization benchmark versus SOTA VLA models; Gemini Robotics-ER achieves 2–3x success rate over Gemini 2.0 on end-to-end robot control. Tested across ALOHA 2, Franka, and humanoid Apollo platforms with partners including Boston Dynamics, Agility Robots, and Apptronik. Accompanying release: ASIMOV dataset for embodied AI safety evaluation.

Implications

The robotics thread, from lab to ecosystem. Boston Dynamics, Agility, and Apptronik as named partners is the robotics ecosystem strategy: Google provides the language-reasoning-to-action model; hardware partners provide the embodiment. The model-hardware stack mirrors what OpenAI tried with Figure AI — but Google’s partnering with multiple humanoid vendors rather than one.

VLA models as the new benchmark surface. “More than doubles performance on a comprehensive generalization benchmark” is the headline claim, but the benchmark isn’t named. That’s a gap — until the benchmark is specified, the 2x claim is unverifiable. Watch for the ASIMOV dataset and benchmark details in the technical paper.

Dexterity demos hide the deployment gap. Origami folding and bag packing are impressive dexterity demonstrations, but they’re controlled-environment tasks. The distance from controlled-environment demo to unstructured real-world deployment is where robotics models have historically struggled. The “trusted tester” framing signals Google knows it’s still early.

Watch:

• ASIMOV dataset adoption by robotics AI research community — it’s intended as a shared safety benchmark
• Boston Dynamics product integration timeline — when does Gemini Robotics appear in a commercial Spot or Atlas workflow?
• Gemini Robotics-ER API availability — can third-party robot developers access the extended reasoning layer?