Inside Gemini Robotics: Google DeepMind’s AI for Every Robot

Ten years ago, Google DeepMind’s robotics team believed a single model could learn to go from pixels to motor control. People thought the idea was crazy. Now, many of the same ideas are reshaping general-purpose robotics.

In this episode of Automated, Brian Heater speaks with Carolina Parada, VP and head of robotics at Google DeepMind, about the company’s Gemini Robotics models and its plan to build an AI layer that can work across many different robot bodies, tasks, and environments.

Carolina explains why robots face an even harder challenge than self-driving cars. Roads have lanes, signals, and established rules. General-purpose robots must interact with people, adapt to unfamiliar spaces, and understand how to “read the room.”

She breaks down how Gemini Robotics combines embodied reasoning, vision-language-action models, and whole-body control. She also explains why on-device AI can be more reliable when connectivity is limited, even when a smaller local model is less capable than one running in the cloud.

Brian and Carolina also discuss why useful robotics does not have to wait for physical AGI, how Gemini Robotics is expanding what Boston Dynamics’ Spot can do during inspections, and why logistics and manufacturing are likely to see capable systems first.

The conversation explores Google DeepMind’s partnerships with Apptronik, Boston Dynamics, and Agility Robotics, the need to evolve AI and hardware together, and the importance of training robots on diverse data. Carolina explains why robot data alone cannot scale and why the path forward will include simulation, human video, and robots learning from their own mistakes.

If you want a grounded look at the models, data, and deployment strategy behind one of the most ambitious efforts in physical AI, this is the conversation.

KEY MOMENTS
00:00 Google DeepMind’s bet on one AI layer for many robots
01:28 Meet Carolina Parada, VP and head of robotics at Google DeepMind
02:57 Advanced Vision & AI Conference
03:45 How cartoons inspired a career in robotics
05:21 Why speech recognition became her machine learning training ground
08:02 Turning speech research into Google products
09:30 Moving from speech to self-driving at NVIDIA
10:05 Finding technical fields ready to be disrupted
11:43 What it means for an AI system to be “solved”
12:12 Why general-purpose robots must read the room
14:28 Why the safety bar must be higher than human performance
16:26 Rewriting systems instead of making incremental improvements
18:35 The single-model idea people thought was crazy
19:06 How reinforcement learning changed whole-body robot control
20:25 Building teams that challenge assumptions
22:36 The debate over one foundation model versus multiple systems
24:00 The three layers behind Gemini Robotics
26:56 How embodied reasoning orchestrates a robot’s actions
27:32 Natural language and the future of multimodal robot control
29:04 Why hard-coded automation cannot handle a changing world
30:26 What robots need to become genuinely useful
32:54 Why one robot will not win it all
33:41 Creating value before physical AGI
34:03 How Gemini Robotics expands Boston Dynamics Spot inspections
35:52 Adding reasoning to robots already in the field
36:55 How the robotics team works across Google DeepMind
38:26 What Google learned from its earlier robotics bets
39:38 Using real applications as benchmarks for broad models
41:16 Why Apptronik became an early Gemini Robotics partner
44:42 Evolving robot hardware and AI together
46:25 How limited training data teaches robots the wrong lesson
49:27 Why robot data alone cannot scale
50:54 Automated’s upcoming live recording and happy hour

Connect with Carolina Parada
https://www.linkedin.com/in/carolinaparada

Explore Gemini Robotics 2
https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/

Register for the Automated Happy Hour with Rodney Brooks
https://luma.com/n688dcl5

Learn more about the Advanced Vision & AI Conference
https://www.automate.org/events/advanced-vision-and-ai-conference

We’d love to hear from you.
Have thoughts or guest suggestions? Reach us at podcast@automate.org

You can find the transcript and more episodes of Automated at automated.fm

Unlock full access to Automated and explore everything automation.
Subscribe today and leave a review on YouTube, Apple Podcasts, Spotify, and the Automated Newsletter.
https://www.youtube.com/@automatedpodcast
https://podcasts.apple.com/us/podcast/automated-with-brian-heater/id1837762221
https://open.spotify.com/show/60olq6brlBEIJWggx2fMR6
https://www.automate.org/automation/automated-newsletter

You can also find us on:
LinkedIn: https://www.linkedin.com/showcase/automated-podcast-by-a3/
Instagram: https://www.instagram.com/automatedpod/