A specialized robot is built to do one task very well, while a generalist robot, also called a general purpose robot, runs one model that can learn many tasks. People who work in software already know why that difference matters, because it is the same difference as the one between a script and an LLM.
Why is an LLM useful if tools and scripts can automate everything?
A script is just a tool that someone built for one job, like a spreadsheet formula, a macro, or a small program. Tools and scripts cannot automate everything. They can only automate what someone already sat down and built a tool for. Every tool needs a builder, a clearly defined input, and enough repetition to be worth the effort.
So most work never got a tool. Think of the email you write once, the messy spreadsheet a colleague sends you, or the report that changes every week. None of it was impossible to automate. Building the tool just never paid off.
An LLM changes that math, because you can simply say what you want in plain words. There is no builder and no setup, and it can handle inputs it has never seen before. It does not replace tools, but it covers all the work that tools were never worth building for, which turns out to be most of the work.
How does the same idea apply to robots?
A specialized robot works like a tool. Someone engineers it for one task, such as an arm that welds the same car door or a machine that fills the same vial. It is fast and reliable, but it is only worth building when the task repeats millions of times. I will call this kind of automation Robotics 1.0: one machine, engineered for one task, in a space arranged around it.
A generalist robot works like an LLM. It runs on a robot foundation model, one model trained on demonstrations of many different tasks, so you can show it or tell it a new task and it does that task in a space it was not built for. It does not replace the welding arm, but it can take on everything that was never worth building a dedicated robot for, like a factory that makes 200 different products, a kitchen, or a warehouse shelf that changes every day. I will call this Robotics 2.0: one model that learns many tasks and works in places as they already are.
What is the difference between generalist and specialized robots?
A specialized robot, sometimes called a specialist robot, is built for one task, in an environment arranged around it. The factory puts every part in an exact position so the robot never has to figure anything out.
A generalist robot is one model that can be instructed to do many tasks, in environments as they already are, so it has to deal with the mess on its own.
| Specialized robot | Generalist robot | |
|---|---|---|
| Era | Robotics 1.0 | Robotics 2.0 |
| In software terms | A tool or script | An LLM |
| Built for | One task | Many tasks |
| Environment | Arranged around the robot | Taken as it is |
| To change the task | Engineers rebuild the setup | Someone gives a new instruction or demonstration |
| Worth it when | The task repeats millions of times | The work varies too much to engineer |
What are real examples of each?
Specialized robots are already everywhere, even if you rarely see them. Welding arms in car factories, machines that place chips on circuit boards, liquid handlers in drug labs, Amazon's warehouse drive units, and Miso's burger flipping robot all belong here.
Generalist robots are much newer. Physical Intelligence's π models fold laundry and clear tables. Google DeepMind's Gemini Robotics follows spoken instructions on several kinds of robots. The humanoids from Figure and Tesla are bets on the same idea.
How are the economics of Robotics 2.0 different from Robotics 1.0?
In Robotics 1.0, almost all the cost is engineering, and you pay it again for every task. Each cell needs custom fixtures, safety guarding, and programming, and installing one often costs two to five times the price of the arm itself. That money only comes back if the cell runs the same task for years. Changing the task means paying engineers again, which is why 1.0 lives in car plants and electronics lines with huge, stable volume.
In Robotics 2.0, most of the cost moves into the model. A company trains it once on a large amount of real robot data, and every customer and every task shares that cost. Adding a new task costs a few demonstrations or an instruction instead of a new cell. That makes it worth automating work that happens a hundred times a week, not only work that happens a million times a year.
The new costs are different in kind. The model needs a steady supply of real world data to keep getting better, and it still fails on situations it has not seen, so someone has to watch it and step in. In 1.0 you pay for engineering up front. In 2.0 you pay for data and for the failures the model cannot handle yet.
Will generalist robots replace specialized robots?
Probably not, in the same way that LLMs did not replace tools. The welding arm will keep welding, because nothing beats a machine built for a task that repeats millions of times.
The more likely future is the two working side by side, and a drug lab is a good example. A liquid handler, the specialized machine, pipettes thousands of samples a day with perfect precision, but only if the tubes are already sitting in its rack. Today a person unpacks the boxes, sorts the tubes, and loads them. A generalist robot can take over that messy part, because the boxes arrive differently every time, while the liquid handler keeps doing the precise, repetitive work.
Software already shows a second pattern. An LLM writes a script, and then the script runs for years. Robotics may end up in the same place, with a generalist model working out a new task and a cheaper specialized routine running it on the line.
The short version
Tools and scripts automated the repetitive work, and LLMs automated the rest. Robotics 1.0 automated the factory line with specialized machines, and Robotics 2.0 goes after everything else with generalist ones. In many places the two will work together, with generalist robots preparing messy work for specialized machines to finish.