A
Alexandra Belski
Guest
As we build humanoid robots to resemble us, it’s easy to imagine that they think with their heads, too. We picture a kind of brain inside that metal head, doing something like what our own brains do. But the reality can be very different. When you talk to a robot half a meter away, your words might actually travel halfway around the world before the robot is ready to respond. Read on to find out where a robot’s thinking actually happens, why sending information across the globe can sometimes be the fastest choice, how robots can share a brain, and more.
A Brain on the Edge
Imagine a robot reaching for a cup that suddenly slips. It needs to react immediately, without sending information to a data center and waiting for an answer. That's where edge AI comes in: processing information close to where it's collected, either inside the robot or on a nearby computer. Keeping that processing nearby helps the robot respond quickly.
But even within a robot's body, thinking can be distributed. A robot can divide tasks between a central system and smaller controllers closer to its moving parts. Think of it this way: one computer decides where the robot's arm should go, while a smaller controller adjusts the fingers' grip as they touch an object. Different parts can handle different jobs, each at the speed required.
When to Put the World on Pause and When Not To
Processing information close to the robot helps reduce delays because data doesn't have to travel to a server and back. But speed alone isn't enough. A robot must also respond at the right moment. A cup can slip, an intruder can step into the path, or a robot can start to wobble. In each case, fast reaction is crucial.
Different decisions may require different planning. Imagine a robot serving drinks. Deciding which cup to pick up can take a little longer, while adjusting its grip to keep the cup from slipping requires a fast response. The robot may use different processes for these tasks.
Besides reaction speed, consistency matters, too. Engineers use the term jitter to describe variations in timing. A robot might be designed to handle a delay of a certain length, but unpredictable delays are much harder to manage. If a response arrives quickly most of the time but occasionally comes too late, that can still be a serious problem. A robot's systems must be designed around which decisions can wait and which cannot.
Farther Away Can Sometimes Be Faster
You might think the best solution is to keep all of a robot’s computing inside its body. After all, the closer the computer, the less time information takes to travel. But there’s a catch: a robot’s onboard computer may be much less powerful than one in a data center.
When a robot sends a task to a remote computer, the information has to travel there, and the answer has to travel back. That takes time. On the other hand, a powerful computer might find a solution faster, so the robot still gets its answer sooner, even if travel is included. Dreamscale Labs argues that as AI models advance, robots will increasingly rely on the cloud, while smaller onboard systems handle essential tasks.
Robots face practical limits on space, battery power, and computing capacity. Designers must decide what a robot needs to process locally, what can be sent elsewhere, and how to balance computing power against the time it takes to get an answer.
Sometimes, the fastest way for a robot to think is to let a more powerful computer do the work, even if that computer is far away.
One Brain, Many Robots
Imagine ten robots picking up cups. You might picture ten separate bodies, each with its own brain. But that's not necessarily how it works. Several robots can use the same remote AI service, sharing access to computing resources instead of each needing its own computer.
Google DeepMind's RT-2 research provides an example of robots accessing AI models through a server, proving that a robot's intelligence doesn't have to come entirely from computers inside its body.
That doesn't mean the robots automatically share memories, learn from one another, or behave like a collective mind. At least, not yet.
The Cost of Thinking
But computing power comes at a cost, and for robots, that cost includes energy.
Onboard computers draw power from the same battery that runs the robot's other electronics and motors. More powerful hardware can also add weight and create extra heat that needs to be managed. In experiments involving the Stretch-3 robot, Microsoft reported that moving AI processing to a remote computer extended battery life by more than 100% in some tested configurations.
That result applies to specific experiments, not every robot. Moving computing elsewhere doesn't eliminate its energy use; it shifts some of that demand to another computer. The robot still needs power to communicate, and the remote server needs electricity too.
Designers therefore face a balancing act: how much computing should a robot carry, and how much should it borrow from elsewhere? The answer depends not only on the task but also on the robot's size, battery, and working environment.