The Head of Robotics at Google DeepMind Discusses General Purpose Robots, Generative AI, and Office WiFi

Earlier this month, Google’s DeepMind team introduced Open X-Embodiment, a database of robotics capabilities developed in collaboration with 33 research institutes. The researchers involved likened this system to ImageNet, a significant database established in 2009, now housing over 14 million images.

The Head of Robotics at Google DeepMind Discusses General Purpose Robots, Generative AI, and Office WiFi
The Head of Robotics at Google DeepMind Discusses General Purpose Robots, Generative AI, and Office WiFi

The researchers, Quan Vuong and Pannag Sanketi, stated that just as ImageNet advanced computer vision research, they believe Open X-Embodiment can have a similar impact on robotics. They emphasized that creating a diverse dataset of robot demonstrations is a crucial step in training a versatile model capable of controlling various robot types, following varied instructions, performing basic reasoning for complex tasks, and generalizing effectively.

The Head of Robotics at Google DeepMind Discusses General Purpose Robots, Generative AI, and Office WiFi

At the time of its announcement, Open X-Embodiment featured over 500 skills and 150,000 tasks collected from 22 robot embodiments. While not reaching ImageNet’s scale, it’s a promising beginning. DeepMind then used this data to train its RT-1-X model, which, when applied to robots in other labs, achieved a 50% success rate, surpassing the in-house methods developed by those teams.

I’ve emphasized this point repeatedly on these pages, but it’s undeniably an exciting era for robotic learning. I’ve had discussions with numerous teams approaching the challenge from various perspectives, each demonstrating increasing effectiveness. While custom-made robots still have their place, it appears that we are gradually moving towards a world where general-purpose robots could become a reality.

Simulation will certainly play a significant role, alongside AI, including the generative kind. It seems that some companies have prioritized hardware development for general tasks prematurely, but in a few years, the landscape might look different.

I’ve been attempting to schedule a conversation with Vincent Vanhoucke for a while. Whenever I was free, he wasn’t available, and vice versa. Fortunately, we were able to coordinate our schedules towards the end of last week.

Vanhoucke took on the position of head of robotics at Google DeepMind in May, marking his recent transition. However, he has been part of the company for over 16 years, with his most recent role as a distinguished scientist for Google AI Robotics. In all, he is likely the ideal person to discuss Google’s robotic aspirations and its journey to this point.

When Did the Robotics Team at DeepMind Come Into Existence

Initially, I was part of the Google Research team and not on the DeepMind side. Our recent merger with DeepMind is what brings my current involvement. However, it’s important to note that robotics research has a more extended history within Google DeepMind. This initiative began when the perception technology was advancing significantly.

Many areas such as computer vision and audio processing were making significant strides, approaching human-level capabilities. This led us to question the potential outcomes if this progress continued in the coming years. One clear implication was that the use of robotics in real-world settings would become feasible. The ability to adapt and perform tasks in everyday environments hinged on having exceptionally robust perception. Initially, I focused on general AI and computer vision, and I had previous experience with speech recognition. Recognizing the changing landscape, I decided to shift our research focus towards robotics as the next phase of our work.

As far as I know, many members of the Everyday Robots team joined this new team. Google has a long history in the field of robotics, going back a decade to its acquisitions, such as Boston Dynamics. It appears that numerous individuals from those acquired companies have become part of Google’s existing robotics team.

Check These Out 


Please enter your comment!
Please enter your name here