Introducing Gemini Robotics 2

Summary
In this episode of Google AI's Release Notes, host Logan Kilpatrick sits down with Google DeepMind robotics leaders Carolina Parada, Stuart Bowers, Kanishka Rao, and Jie Tan to discuss the announcement of Gemini Robotics 2. The panel covers advances in whole-body control, dexterous manipulation, multi-robot collaboration, and the release of Gemini Embodied Reasoning (ER) models and Vision-Language-Action (VLA) models.

What is shown

• Roundtable Discussion [00:38]: Logan Kilpatrick discusses robotics timelines and technical hurdles with the Google DeepMind robotics team.
• Lamp Switch Flip [28:16]: An Apollo humanoid robot autonomously flips a toggle switch on a lamp using its robotic hand.
• Kitchen Cleanup / Sweeping [28:31]: A humanoid robot uses a hand broom and dustpan to sweep debris off a counter in real-time autonomous operation (1x).
• Ziploc Packing [28:44]: Robot hands autonomously place grapes into a plastic Ziploc bag and manipulate the seal to close it.
• Lightbulb Removal [29:01]: Robot Apollo unscrews a lightbulb from an adjustable desk lamp using multi-fingered coordination.
• Tool Kit Organization [29:25]: A Franka arm equipped with a parallel gripper picks up tools (such as a hammer) and precisely slots them into a molded plastic toolbox.
• Trash Bag Knot Tying [29:47]: A humanoid robot coordinates two multi-fingered hands to loop and tie knots in a trash bag drawstring.

Claims & numbers

• Jie Tan states that 3 years ago he thought robots in daily life were beyond his lifetime, 2 years ago he estimated 10 years, and currently estimates 5 to 10 years [00:12, 04:33].
• Carolina Parada claims Gemini Robotics 2 brings whole-body intelligence across multiple robot form factors, enabling reasoning over complex multi-step spatial tasks and multi-robot collaboration [02:20, 04:27].
• Jie Tan notes that human hands have over 20 degrees of freedom, making contact-rich dexterous manipulation vastly harder than locomotion [10:24].
• Kanishka Rao and Jie Tan describe the robotics data pyramid from costly teleoperation down to wearable grippers (like UMI) and egocentric human video [12:20].
• Kanishka Rao notes that the robotic hands shown on the GR2 humanoid have 20 independent degrees of freedom/joints per hand [27:58].
• Carolina Parada notes that roughly 90% of an organization task is semantic/spatial reasoning, while the remaining 10% is exact low-level physical execution [22:13].
• Stuart Bowers states that the Embodied Reasoning 2 model is being released directly via Google AI Studio and API, alongside an on-device action model for trusted testers [35:20].
• Carolina Parada announces an open-source safety evaluation benchmark called "Asimov" for agentic physical decision-making [38:16].

Notable quotes

• "What we're building here is the intelligence layer to power any robot to do a broad range of useful tasks." — Carolina Parada [01:16]
• "Locomotion is nearly a solved problem. What's remaining, it's actually a very hard problem, is dexterous manipulation." — Jie Tan [09:57]
• "On the real robot, it's very difficult to write deterministic code that can actually take in just a raw set of pixels from a bunch of different cameras and actually give you thoughtful, correct joint angles." — Stuart Bowers [33:20]

Assessment
This is an official Google DeepMind launch discussion and demonstration video for Gemini Robotics 2. While the discussion outlines high-level concepts and long-term deployment hurdles candidly, the video clip demonstrations are tightly curated highlight reels of dexterous tasks shown at autonomous 1x speed.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.