Picking from a moving conveyor with imitation learning
In a master thesis at TOS Robotics, Nienke Dekkema replaced the ACT model of the earlier TOS Robotics generation with a model that learns its image features from the setup itself and has memory over time. In a controlled conveyor setup with one test object, success rose from 18% to 98% at 0.10 m/s and from 0% to 90% at 0.15 m/s.
Problem
On a moving belt a small timing error can mean a missed grasp. The baseline was the model of the earlier TOS Robotics generation, a ResNet-34 image network with ACT (Action Chunking with Transformers). According to the thesis it could pick a bell pepper from a belt that stood still, but not from a moving one, even at constant speed. The question was how to make it grasp reliably when the belt speed varies between trials.
Setup
- an Interbotix robot arm with a gripper, driven at 100 commands per second within limits for speed, acceleration and jerk
- the TOS Robotics Teachbot for teleoperation
- two colour and depth cameras at 90 frames per second, one fixed, viewing the belt, and one on the wrist
- a small red cylinder at a constant belt speed per trial, 0.10 or 0.15 m/s, which the robot grasps, carries back to the start of the belt and releases
The data held 75 trials per speed, 50 for training and 25 for validation, made by replaying one teleoperated demonstration with speed-dependent timing offsets. Each model was tested 100 times per speed. Reliable means at least 90% success.
Results
| Model | 0.10 m/s | 0.15 m/s |
|---|---|---|
| ResNet-34 with ACT, blended chunks of 75 actions | 18% | 0% |
| ResNet-34 with LSTM, memory over two images 40 ms apart, one joint target 80 ms ahead every 40 ms | 65% | 0% |
| Variational autoencoder with LSTM, a small image model trained on cropped belt images and the cylinder position | 98% | 90% |
The baseline started too late or too early, at 0.15 m/s too late in 99 of 100 trials. The author suggests that ACT without memory cannot tell the two speeds apart and that chunking reduces reactivity. In an extra test at one speed, 0.15 m/s with 30 trials per setting, the blending weight decided the result: success rose from 30% with the newest prediction weighted at 0.95 to 100% at 0.3 or lower. Trained on several speeds, ACT was still not reliable.
The ResNet-34 with LSTM model did well on validation data at both speeds, and the author reads the gap on the robot as ResNet-34 features generalising poorly to new images. The variational autoencoder with LSTM was reliable at both speeds and could recover after starting too early. In 83% of its failed trials the arm made sudden, large movements, which the thesis links to missing camera frames that the system replaced with a black image.
Limits
The test used one cylinder at two speeds, a dataset built from one demonstration and a pass or fail score. Missing camera frames were not addressed. Food products still have to be tested. How TOS Robotics smooths its own robot movements is described in smooth robot motion with ACT.
Full title: End-to-end imitation learning for conveyor pick-and-place tasks under speed variation. Carried out at BIASlab in collaboration with TOS Robotics.
Questions and answers
Which change made the difference?
The thesis cannot separate them. The variational autoencoder model differs from the ResNet-34 with LSTM model both in the image model and in the input image, which was cropped and smaller, so the gain cannot be attributed to one of the two.
Is this used on the production line?
No. The production cells of TOS Robotics use a point cloud diffusion policy with a separate path planner and follow the belt with line tracking. Applying these findings there would be new work.