Why demonstration data needs cleaning before a robot learns from it

A robot that learns from demonstrations also learns from the bad ones, so TOS Robotics cleans the data before training. We learned that cleaning can go wrong too, because our first automatic filter threw away about 30% of good data. In 2026 we moved from lists kept by hand to limits based on what the robot can do, and to a human review of failed picks from the line.

First setup

Cleaning started at the source. While a TOS engineer records demonstrations with the Teachbot, one command marks the last recording as a mistake. In the datasets we recorded between December 2025 and March 2026, about 1 in 13 demonstrations got that flag, from 0.8% to 14.7% per dataset. The rest was hand work. Each training script carried a typed list of samples to skip, about 500 entries over four datasets for one deployed model.

What went wrong and what we changed

In March 2026 we wrote an automatic cleaner and later scored it, check by check, against the human lists.

Problem What we measured What we changed
Outlier filter that cut the most extreme few percent of each measurement About 30% of clean samples rejected, although the extremes are real variation Limits from the extremes of 803 successful production cycles plus 5 mm per side, at most 1.3% of good picks rejected
Two other checks Fewer than one in four of their rejects were on the human lists Both checks switched off
Missing camera calibration, the numbers that place each camera relative to the robot 786 of 1,636 cleaned samples Fixed calibration added
Different depth offsets of the two cameras Point clouds about 20 mm apart Offsets compensated
Poses the robot’s guards, its joint limits and safety zones, would block About 3% of demonstrations and about 24% of human corrections Filtered out, and the review tool blocks accepting them

Relabelling failures

When the deployed model fails a pick on the line, someone from our team opens that cycle in 3D and moves the pick to where it should have gone. The correction goes into training only, never into the test set. This method is called DAgger, short for Dataset Aggregation. Our tool first proposed corrections itself and skipped failures it judged uncorrectable. We removed both, and an unchanged pick cannot be accepted.

Not every failure is a label problem. On one day about a third of the failed picks were two peppers lying on top of each other. A small share of cycles, where the pepper moved after the picture, held about a third of all failed picks. About half of the reviewed failures showed no visible error in the pick position. Relabelling those teaches the model noise. A blind second labelling of one label class also found that 27% of the original labels were over-calls, because the labeller’s criterion had drifted.

Results

A few dozen corrected failures gave mixed offline results, with the position error slightly worse, within the noise between training runs, and the rotation error better. The clearest measured gain came from scale. We then trained on 102,973 production cycles with audited outcomes, 11,130 of them with a corrected target. In an offline test on a held-out day, place flips of more than 90 degrees fell from 5.7% to 2.1% against the previous model. That model went live on both robots in September 2026.

Questions and answers

What is demonstration data?

Recordings of a person doing the task with the robot. For each moment the system stores what the cameras saw and how the robot moved, and the AI model learns from these examples.

Which cleaning ideas did not work?

We tried to predict the human accept and reject decisions from the pick geometry. It did no better than a coin toss, and we keep it documented so nobody repeats it.

See it run on your product

Send us your product and your packing format. We show you what the robot does with it, in simulation or on our test cell in Vught.

Request a demo