Robot technology is moving forward, thanks to heroic efforts by Boston Dynamics and various robot vacuum companies. But it seems like a gap is opening up between what chatbots like Gemini, Claude an ChatGPT can do, and the capabilities of physical machines. The former are writing symphonies while the latter are struggling to stumble across an empty room.
Incredibly, Google engineers foresaw this problem nearly a decade ago. Boffins believed that the invention of the transformer would assist with artificial cognition and improve language models, but they didn’t think it would translate easily into physical activity.
“This problem is currently dogging the household appliance sector,” says Vacuum Wars, a company that is always looking for the best robot vacuum on the market.
“Getting these devices to sense their environment is becoming easier with camera and LiDAR technology, but giving them appendages that let them navigate it is far more challenging.”
That’s where Google comes into the picture. One of its “moonshots” is to give robots an AI body. The hope is that infusing them with intelligence will provide them with a newfound ability to get out of silicon and into the real world, which, for many, would be the holy grail.
An army of ten million robots could replace an entire country’s physical labor force and provide additional services at the margin, like cleaning streets, clearing trash from the oceans, and helping little old ladies across the road.
You could see widespread automation across the transport system and even new bot-like devices for taking people from A to B.
Google’s aim to give AI a robot body began in earnest in 2017 when most of the investor community hadn’t heard of artificial intelligence, or if they had, just said it was useless, like Siri.
The company wanted to improve the success rates of its robots using grippers to pull, push and manipulate objects. The thinking was to build software and hardware that could emulate the human hand, nature’s most prolific tool-building device.
Initially, the success rate was around 7 percent. Robot arms could sometimes grip and manipulate objects in their vicinity, but they couldn’t do it reliably. It was a proof of concept, but little more than that.
However, the robot arms got better through AI-based training. Eventually, they began to learn which tactics worked for specific shapes through brute force trial-and-error methods. Their processes didn’t quite mimic human learning, but it was clear that they were getting the basics right.
Eventually, Google’s robots could reach into a tray, follow instructions to grab a yellow LEGO brick, nudge the others out of the way, and place it next to another block while following human instructions. The breakthrough came when Google transitioned to using heuristics instead of hard-and-fast rules.
The idea was to give the robot more leeway in how it learned, offering it the chance to experiment with various options and ultimately land on a method that might work (such as deliberately moving other blocks out of the way first before attempting a grab operation).
The other big innovation was getting the robots to perform the maneuvers in simulations.
Instead of waiting for the robot arm servos and motors to fire, pick up an object, and then learn from their mistakes, Google got the equipment to run through millions of simulated trials of what might happen first before selecting a candidate methodology to attempt in the real world.
Again, this method sped up the learning process and could be applied to numerous operations, including walking, or standing tall.
“We are currently seeing some of the fruits of this technology in the robot vacuum space,” says Vacuum Wars.
“There are hints that the dumb older technology is giving way to more complex machines that seem to be able to do whatever consumers want. Robot vacuums already have legs. It’s not long before they also develop arms and can do the washing up.”
The technology is also the sort of thing that keeps its developers up at night. Google’s head of AI-powered robotics, Hans Peter Brondmo, said that the prospect was alluring to him because it means that human physical limits would finally be breached.
The amount of computation required to get robots to pick up LEGO blocks was enormous. Google had to commit entire server farms to it.
But once a machine learns a new skill, it could immediately share it with thousands of other robots, improving their capabilities in an instant without the need for further education.
This instant transmission means that robots are always on the frontier of what the technology makes possible and can receive updates in real time. Any innovation that works and fulfills the robots’ objective functions is potentially a way to boost outcomes and achieve Brondmo’s dream.
“One day, everyone woke up to ChatGPT and saw that the world had changed,” says Vacuum Wars.
“The same thing will probably happen in the robotics space as the exponentials converge. One moment, robot vacuums are crawling across the floor on wheels, the next they are whipping up strawberry blancmange in the kitchen and cleaning the toilet.”
However, experts like Brondmo suggest people temper their expectations to a degree. While robots playing football better than Real Madrid by 2050 is a possibility, he suggests that there might be hard limits on what AI can do.
Some complex tasks in the real world may still require human direction, even if robots develop the dexterity to perform all manual tasks.
Systems might not have the agency or initiative to go to the next step and think for themselves what’s sensible to do next, and what makes sense.
For example, they might be able to make a bet, but whether they will know how to move between hotel rooms and when it is appropriate to clean a room is another matter.
Furthermore, a core tenet of robotics safety is that devices shouldn’t have agency. Robots must respond to human commands instead of following their desires. Perhaps that’s why Google hasn’t released anything yet and may not for a while to come.
Main image courtesy of Unsplash
