# Research and lessons from building AI packing robots

TOS Robotics does research with Eindhoven University of Technology on robots that learn tasks from demonstration. This page collects what we learned on a real packing line, the master theses done at TOS Robotics and our funded projects.

Source: https://tosrobotics.com/research/
Last updated: 2026-11-14

## Lessons from the line

- [Why we keep USB out of our robot cells](https://tosrobotics.com/research/why-we-avoid-usb-in-robot-cells/): A USB depth camera needs about 19 seconds to start again. We moved the encoders and inspection cameras to Ethernet or a fieldbus.
- [How do you make ACT move a robot arm smoothly?](https://tosrobotics.com/research/smooth-robot-motion-with-act/): ACT took 17.1 ms per prediction on average on our robot and 97.9 ms at worst. The chunk handover is sized for the slow case.
- [Why a taught robot path cannot simply be sped up](https://tosrobotics.com/research/why-taught-robot-paths-cannot-be-sped-up/): Raise the speed and a blended corner moves. On our robots the same CNT50 corner cut 14 mm from a standstill and 24 mm from a moving start.
- [Why we built our own Teachbot](https://tosrobotics.com/research/why-we-built-our-own-teachbot/): Encoders 32 times finer than the off-the-shelf arm we first taught robots with, and joints of 57 g.
- [Why small errors cause big losses on a packing line](https://tosrobotics.com/research/small-errors-big-losses-on-a-packing-line/): On one bad day, the line standing still cost about seven times more packs than the robots' missed and failed picks together.
- [How to build fault screens that operators actually see](https://tosrobotics.com/research/fault-screens-operators-actually-see/): One banner per event instead of up to three, and a photo of the machine with the part to fix outlined in red.
- [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.
- [Why we moved from ACT to a diffusion policy with point clouds](https://tosrobotics.com/research/diffusion-policy-versus-act/): We sample 50 candidate picks instead of one average. In nearly all cycles at least one of them was a working pick.

## Master theses at TOS Robotics

- [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%.
- [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.
- [Imitation learning when there is more than one right way](https://tosrobotics.com/research/imitation-learning-action-multimodality/): Giving ACT its own planned actions as input cut first joint deviation by 5 to 22%, but failed tasks did not go down consistently.

## Research projects

- [ATLAS: robots that learn the choices operators make](https://tosrobotics.com/research/atlas/): In ATLAS, TOS Robotics and TU Eindhoven develop imitation learning that copies the choices of an operator, with multi-camera sensing that self-calibrates.
