Carolina Parada: “AI has to be smart enough to embody any robot.”

In this talk

1. 1:49

   From speech research back to robotics

   Parada traces her path from electrical and computer engineering and control systems to a PhD in speech recognition. After working on Google speech products and self-driving perception at NVIDIA, she returned to Google to lead robotics research.

2. 5:03

   Building AI for many kinds of robots

   Google DeepMind’s robotics mission is to build embodied AI for the physical world. Parada describes language and vision-language models for planning, vision-language-action models that turn concepts into actions, and research spanning locomotion, dexterity, laundry, and in-hand manipulation. The aim is for the same intelligence to work across different robot bodies.

3. 9:31

   Reasoning, acting, and transferring between robots

   Parada says Gemini Robotics 1.5 combines stronger spatial reasoning and tool use with a vision-language-action model that can think while acting. A robot could, for example, look up local recycling rules before sorting trash; the acting model can continue working on a multistep task for 15 minutes. Cross-embodiment transfer uses the same weights across robots such as ALOHA and Apollo, though training still needs data from multiple embodiments.

4. 15:32

   Learning from data and proving generalization

   Teleoperation is currently the strongest source of robotics training data, with simulation the second strongest and also useful for evaluation before real-robot trials. Parada says proving generalization requires testing across objects, scenes, motions, and robots, while broader benchmarks and physical evaluations remain necessary. She expects imitation learning and reinforcement learning to complement each other.

5. 30:30

   Deployment, homes, and responsibility

   Google DeepMind is engaging external users through an Apptronik partnership, a trusted tester program, and access to its reasoning model in AI Studio. Parada sees humanoids as a valuable form factor, while emphasizing that robotics will include many types of robots. Homes may be among the last settings for deployment because of safety, privacy, social acceptance, and policy challenges; she estimates five to ten years for stronger models, while noting that other hurdles remain.