The capabilities behind our products
TOS Robotics builds its products on four capabilities. Robot control, AI algorithms, hardware design, and digital twin and simulation. Our robots do not follow programmed positions. 3D cameras see each product and an AI model, trained from demonstrations, decides how to pick and place it. That is why they work on products that never look the same twice.
Robot control
Line tracking, path planning, motion streaming and safety. The part that moves the arm.
AI algorithms
Imitation learning, a diffusion policy on 3D point clouds and clean data from the line.
Hardware design
Our own robot arm, Teachbot, grippers and complete cells, designed for production.
Digital twin and simulation
A working 3D copy of your line. See it run before you order, and test changes without stopping production.
Which products use them
Robot control
The flowpacking cell, the TOS Robot and the AI robot control software.
AI algorithms
The flowpacking cell, the Teachbot, quality control and the picking and placing AI models.
Hardware design
The flowpacking cell, the TOS Robot and the Teachbot.
Digital twin and simulation
The flowpacking cell and the TOS Robot.
From demonstration to reality
An engineer of TOS Robotics shows the task on real products with the Teachbot.
It learns what a good pick looks like. It does not memorise positions.
It picks and places whether the belt moves or stands still, also on products it has never seen.
After deployment the AI model keeps learning from its own picks with reinforcement learning. The line gets better while it runs.
New layouts, speeds and packs are tried on a 3D copy of the line first.
Measured on our own systems
- 99.1%
- pick accuracy on the bell pepper line, up from 97.6% in the first weeks
- 1day
- to teach a new task in a pilot at a packer
- 0.02s
- to check one camera image with one or more products
Questions and answers
Does the robot keep learning in production?
Yes. After deployment the AI model keeps learning from its own picks in production with reinforcement learning. A new version is tested before it goes to the line. A master thesis at TOS Robotics showed this on a real FANUC robot. On the bell pepper line the pick accuracy rose from 97.6% in the first weeks to 99.1% by weeks seven and eight.
Does TOS Robotics do research?
Yes. In the project ATLAS, with Eindhoven University of Technology, robots learn the choices behind a demonstration, such as which product to take first. Our research page has the master theses and the lessons from the line.