# AI algorithms: imitation learning and 3D vision

The AI of TOS Robotics learns a task from demonstrations, a method called imitation learning. A diffusion policy looks at each product as a 3D point cloud and proposes where to pick and where to place it. We train it on cleaned data from our own production line.

Source: https://tosrobotics.com/capabilities/ai-algorithms/
Last updated: 2026-11-20

- Method: Imitation learning
- Model: Diffusion policy
- Input: 3D point clouds from two depth cameras
- One model trained on: 102,973 production cycles

We do not program pick positions. **We show the robot the task and the model learns it.**

## What we build

- **Imitation learning.** Our engineers show the task with the [Teachbot](https://tosrobotics.com/products/teachbot/) and the model learns from the recordings. The guide [what is imitation learning](https://tosrobotics.com/knowledge/what-is-imitation-learning/) explains the method.

- **Diffusion policy.** We run the model 50 times per pick and choose from the largest group of answers, so two valid orientations never average into a wrong one. See [why we moved from ACT](https://tosrobotics.com/research/diffusion-policy-versus-act/).

- **3D vision.** Two depth cameras turn each product into a point cloud. In our offline test the mean rotation error fell from 26.0 to 14.9 degrees against camera images.

- **Clean data.** Our team corrects failed picks from the line and they go back into training. See [cleaning demonstration data](https://tosrobotics.com/research/cleaning-demonstration-data-for-robot-learning/).

- **Tested on the line.** Since July 2026 every model change runs as a block-randomised test on the line.

## Where it is used

- [Flowpacking cell: robots that fill the flowpacker](https://tosrobotics.com/products/flowpacking-cell/): A complete cell that fills the flowpacker. €150,000 per cell. In production.
- [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.
- [Vision quality control for fruit and vegetables](https://tosrobotics.com/products/quality-control/): Every product checked by camera and AI in 0.02 seconds, without a person at the belt.
- [AI models for picking and placing fresh produce](https://tosrobotics.com/products/picking-and-placing-ai-models/): Where to pick and place each product. 99.1% pick accuracy. Licence for any robot.

## Keep reading

- [What is imitation learning for robots, and when does it fit?](https://tosrobotics.com/knowledge/what-is-imitation-learning/): Teaching a robot by showing it the task, used by TOS Robotics on a bell pepper line at 40 picks per minute per robot.
- [Why demonstration data needs cleaning before a robot learns from it](https://tosrobotics.com/research/cleaning-demonstration-data-for-robot-learning/): About 1 in 13 demonstrations was flagged as bad on the spot. Our first automatic filter threw away about 30% of good data.

## Questions and answers

### What is imitation learning?

A way to teach a robot by showing it examples instead of programming it. The guide [what is imitation learning](https://tosrobotics.com/knowledge/what-is-imitation-learning/) explains how it works.

### Why a diffusion policy and not ACT?

Because ACT gives one average answer. When a pepper can lie two ways, that average is wrong. [Why we moved from ACT to a diffusion policy](https://tosrobotics.com/research/diffusion-policy-versus-act/) has the numbers.

### Is AI also used for quality control?

Yes. [Quality control](https://tosrobotics.com/products/quality-control/) uses an AI model that judges colour, shape, damage and decay from camera images.
