Why Google DeepMind robotics bets on the brain over the body

This article summarizes an episode of Waymo’s video series featuring Carolina Parada of Google DeepMind.

Image credit: Tech in Asia

The race to build robots that can do many tasks is often seen as an effort to put a human-like robot in every home. But according to Carolina Parada, senior director at Google DeepMind, the important work is happening in other areas.

She says the biggest problems for AI in physical robots are not about what they can do, but about the human challenges of safety and trust.

The brain before the body

Instead of building a robot for every task, Parada believes the goal is to create one flexible AI that can learn to run many robots. This changes the industry’s focus from hardware to software.

A model-first approach
Parada explains, “I am definitely more on the [side of building] the generalist model. I’m a believer that we’re going to have many many different form factors, actually, and humanoid is going to be just one.”

Adaptability is the endgame
This method means one AI improvement could help many different machines, from warehouse arms to delivery drones. The AI can be moved from one machine to another.

She continues, “Even if the embodiment can be specialized, we believe the intelligence should be capable of quickly adapting to even different embodiments and then being able to operate in that way. To build something truly general, [the AI] has to be able to handle a very diverse set of applications.”

The commercial reality check

While people like the idea of a robot for the home, the first robots will be sold for other uses.

The delay for domestic robots
The technical problems of doing chores seem possible to solve, but Parada points to other problems that are harder to fix.

“I definitely think the home will not be the first place that [robotics] will make impact. Not necessarily because we can’t solve many different problems in the home soon,” Parada argues, “but mostly because the home requires a different level of safety, privacy, and security that will just take some time to mature.”

The real proving ground

With the home not being the first priority, the first chance is in factories and businesses. These places need robots that can do the difficult jobs that could not be automated before.

She identifies the first markets, stating, “We can start doing quite a bit in many different spaces: logistics, manufacturing, retail, and even then moving into healthcare. When we’re talking about logistics, we’re talking about the hard things in logistics. Everything that is simple in those spaces has actually already been automated.”

The data dilemma

Safety is not a single problem

Beyond the data problem