# What is imitation learning for robots, and when does it fit?

Imitation learning teaches a robot a task by showing it, not by programming it, and it fits tasks where products or positions vary. TOS Robotics uses it on a bell pepper line in production, at 40 picks per minute per robot.

Source: https://tosrobotics.com/knowledge/what-is-imitation-learning/
Last updated: 2026-11-15

## How it works

1. **Demonstrate.** A person controls the robot by hand, for example with the [Teachbot](https://tosrobotics.com/products/teachbot/), and does the task on real products. At TOS Robotics its own engineers do this.
2. **Record.** For each moment the system stores the camera images and the movement of the robot.
3. **Train.** A neural network learns the link between what the cameras see and what the robot does.
4. **Test.** The model is tried on products that were not in the demonstrations.
5. **Run.** On the line the model decides for each new product what the robot does.

## Imitation learning versus programming

A programmed robot repeats positions and rules that an engineer wrote down. A robot taught by imitation learning has learned from examples what to do with what it sees. It can handle variation that nobody wrote a rule for, as long as the examples covered that kind of variation.

That makes it a fit when products differ from each other, as natural products do, when the task depends on a skilled hand, such as careful handling or an even spray pattern, and when products or packs change often, so reprogramming each time is too slow. It fits less when the task is identical every cycle. A fixed pick from a fixed position is simpler and usually cheaper to program once.

## What it needs

The robot copies what it is shown, so the demonstrations must be careful. The examples must hold real variation, with small, large, crooked and odd products. Before a model goes to the line, it is tested on products it has not seen.

## At TOS Robotics

TOS Robotics uses imitation learning on a bell pepper flowpack line at a packer in the Netherlands, where the robots pick peppers that vary in shape and size from a moving belt. The pick accuracy there was 97.6% in the first weeks and 99.1% by weeks seven and eight.

One research direction is the choices behind a demonstration, such as which product to take first, in which order and along which route. TOS Robotics works on this with Eindhoven University of Technology in the research project [ATLAS](https://tosrobotics.com/research/atlas/).

## Keep reading

- [Teachbot: teach a robot by hand](https://tosrobotics.com/products/teachbot/): A teaching tool. Guide it by hand and the robot follows. Sold on its own.
- [How long does it take to teach a robot a new product?](https://tosrobotics.com/knowledge/how-long-to-teach-a-robot/): Programming a new task takes days to weeks. In a TOS Robotics pilot at a packer, a new task was taught in one day.
- [Picking from a moving conveyor with imitation learning](https://tosrobotics.com/research/imitation-learning-moving-conveyor/): In a lab conveyor test, success rose from 18% to 98% at 0.10 m/s and from 0% to 90% at 0.15 m/s.
- [Reinforcement learning on a real FANUC robot, without a simulator](https://tosrobotics.com/research/residual-reinforcement-learning-real-robot/): In seven days on a real FANUC arm, a residual RL layer raised lab task success from 51.1% to a stable 58 to 60%.

## Questions and answers

### What is the difference between imitation learning and reinforcement learning?

In reinforcement learning the robot learns by trial and error with a reward. In imitation learning it learns from demonstrations, without trial and error on the real line. TOS Robotics uses both, imitation learning to teach the task and reinforcement learning after deployment to keep learning from the robot's own picks.

### Can imitation learning handle a belt that keeps moving?

Timing the grasp is the hard part. On the bell pepper line the robots follow the belt with line tracking. In a lab test with one object in a master thesis at TOS Robotics, a model with memory over time raised success on a moving belt from 18% to 98% at 0.10 m/s.
