This Company Is Building Superhuman Robot Manipulation

This company is building superhuman robot manipulation.

Have you ever seen a robot this gentle and dexterous?

In more than a decade of writing about robotics,

I've honestly never seen anything like it.

The robot was developed by Eka,

a startup based here in Cambridge, Massachusetts.

When I visited the company's headquarters recently,

I watched its robots handle various items.

Eka's offices are filled with robot arms, grippers,

and all kinds of objects

that are used to test robot dexterity.

Take for example, Eka's robot grasping

and screwing a light bulb into a socket.

The claw decelerates gently pinching at the bulb,

and when the glass object spills from its grasp,

it chases it along the table like a dog going after a ball.

Once the robot gets the light bulb between its pincers,

its swiftly screws it into a nearby socket,

illuminating its work area.

Encountering Eka's robots reminded me of the first time

I tried talking to ChatGPT.

It feels as if there's a strange

kind of intelligence behind the machine.

Eka's robots already know how to grasp all sorts of things.

I watched one swiftly, but delicately

handle a tiny raspberry.

The kind of skill that could prove

very valuable in the food industry.

When I tried giving Eka's robot my house keys,

it picked them up easily.

When I tried taking them back,

it resisted for a moment, then let go,

turned its attention back to the table

and hunted for something else to pick up.

Eka is not just building a picking robot though.

Their technology could let robots

take on all kinds of work

from manufacturing electronics to handling suitcases.

Eka was founded by Pulkit Agrawal, a professor at MIT,

and Tuomas Haarnoja, an ex-Google

DeepMind robotics researcher.

The pair say they identified a way to teach robots

much more advanced dexterity a couple of years ago.

If they succeed, it could revolutionize

how robots are deployed in many industries.

Eka is taking a fundamentally different path

to a lot of other robot startups.

Instead of having robots learn from human demonstrations,

they want them to learn inside simulations.

Other startups are training robots

using human videos and teleoperated demonstrations

to build so-called vision-language-action models.

These systems have started to show some promising results

with robots learning to do things like fold laundry

and load dishwashers, albeit very slowly.

Agrawal and Haarnoja say their approach

will let robots not only match human dexterity,

but become superhuman.

They also say their approach should be more scalable.

Since there is no need to collect endless videos

or teleoperated demonstrations,

and because Eka's algorithms learn for themselves,

they instinctively know how to recover

when fumbling something.

This kind of thing is difficult for other robots to learn

unless the humans training them

deliberately make a wide range of mistakes.

Eka's approach may also lend itself to fine dexterity.

This could mean robots taking on

a much wider range of human work,

including things like assembling electronics.

Eka's demo suggest that the company

may be onto something big.

Perhaps their algorithm is similar to OpenAI's GPT-1,

which hinted at what large language models

could eventually become.

Read my full experience at Eka's lab at wired.com.