How To Build A Robot From Nothing
# # ai-research
# # ai-safety
# Robotics
# Senior-expert
Around 3 Years To Build, But How?
October 22, 2026
Leela Krishna
Aleksandar Scekic
How Are Robots Built? The Four Steps To A Prototype
A bare-bones robot costs 30 to 50 thousand dollars, has to sit down for an hour after every hour of work, and still cannot make the same coffee ten times in a row. None of that is the software's fault. So how are robots built, and where does the trouble start?
This comes from our session with Leela Krishna. She leads physical AI research at Centific, our partner: the meeting point of computer vision and robotics. Before that she spent about eight years building autonomous driving and machine learning systems.
Tuesday's post covered what physical AI is. This one answers the question an attendee put to Leela live: if we wanted to build a robot from nothing, what are the steps?
How Are Robots Built? The Short Answer
Robots are built in four steps over two to three years. First, pick the body: four legs, two legs or wheels, matched to the work. Second, pick the brain, a model that is still far from finished. Third, decide what the robot is for, industrial or household. Fourth, teach it, which is mostly data.
Step 1: Pick The Body Before Anything Else
The first decision is anatomy, and it is made before a single line of code. A quadruped walks on four legs like a dog. A bipedal humanoid walks on two. A third group asks why you need legs at all. Legs spend most of their effort balancing the body, so this group puts a torso with arms on wheels instead.
Leela's rule is to decide the anatomy and the application together, because the hardware has to be resilient for the exact work it will do. A warehouse robot that lifts 30 kilos ten times an hour and a kitchen robot that handles 100 light objects are different machines from the frame up.
Step 2: Pick The Brain, Knowing None Is Finished
While the hardware is being engineered, the team looks for a model that can act as the robot's brain, and here the honest news is that the shelf is nearly empty. Language models have mature players. Robot brains do not yet.
"We are still in the GPT-2 era, which means there are no evolved players like ChatGPT in the robot brain world. There are very few models that can do very definite things, and general-purpose models that are still improving."
The team chooses between a narrow model that does a few things reliably and a general model that is still maturing. Steps 1 and 2 together are the first 12 to 18 months of any robotics company, and they end with a hardware prototype ready to go on an assembly line.
Step 3: Decide What The Robot Is For
The next question sounds obvious and shapes everything that follows: what will the robot actually do, and where? Industrial work means heavy lifting, hot parts and the same movement repeated all day. Household work is the opposite, enormous variety and very low complexity per task.
That split decides how you teach it. Leela's point is that the two environments fail in different ways, so you plan both the hardware and the teaching around one of them. A robot built for everything is the humanoid promise from the videos, and it is also why those robots are 10 to 15 years from doing what the videos imply.
Step 4: Teach It, Which Is Where The Years Go
Teaching takes another one to two years, and almost all of it is data. Start with one ability, handling a hot object. The robot needs to feel temperature, which means sensors on the hands. Humans have that already: it is called skin. Now repeat that for every ability a person uses in a day. Most of that data is recorded human demonstrations: a person does the task while a camera and sensors capture every move.
"Take building ChatGPT, multiply it by all the abilities humans have, not just reading and writing. That's the complex task, and most of it is essentially data."
The training data is gathered and annotated the way it is for vision or audio, only with a reasoning layer on top. If a robot is making a pizza and you say "stop", it has to hear you while it works and understand that the instruction is about the pizza. That is multimodal reasoning, and Leela says the field is only starting to see it. The total, in her words, is two to three years of effort to see a bare prototype.
Why Are Robots So Expensive?
Robots are so expensive because the hardware, not the software, is still the costly part. Leela puts a bare-bones robot at 30 to 50 thousand US dollars and a fully taught one around 150 thousand. The model is already cheap to run. Coordinated, resilient hardware is not, and neither is servicing it the way the car industry services cars.
There is a second cost, and it is the one that stops a robot working a full shift. After roughly an hour of work, a robot's motors overheat and it has to be unplugged and left to cool. Cars solved this years ago with liquid-cooled motors, and your laptop has a fan. Robots do not have that resilience yet. Leela's comparison is the iPhone: the first one launched in 2007 with no GPS at all, and it took 10 to 15 years of iteration to become convenient. Robots are at 2007.
Why Robots Fail Where Humans Do Not
Everything in the four steps teaches the robot what engineers call the happy path: the right way to make a good coffee. Failure is still, in Leela's words, an unexplored domain. When you drop a plate you pick up the pieces, throw them away and wipe the floor, and if something slips you can catch it mid-air. Nobody has taught a robot that recovery, and a kitchen produces it daily.
Then there is consistency. Leela makes a coffee she likes every day, at the same standard. Ask a robot to make ten identical coffees for ten guests and it cannot. Someone recently demonstrated a robot making ice cream, and once you count the flavours, it is a 55-step process. Between coffee and ice cream the number of steps jumps, and every step is a new place to fail.
"The robots that we have, they will help us with some tasks in the kitchen. But we would have to help them."
That last line is the whole post in one sentence. Four steps, two to three years, and the result is a machine that still needs people: to build it, to teach it fifty times over, and to clean up when it spills. The next post in this series covers that teaching work and what it looks like for an expert with a phone and no engineering background.
Find a project that knows your worth.
Comments (0)
Popular
Dive in