Library robot: picking and stacking books.
In the trials shown here, a plain ACT policy picked up the demonstrated book, a few books that were not in the demonstrations, and a book that was partially occluded.
Why books
Books are rigid, but they vary in size, thickness, cover design and how they sit on a surface. That makes them an interesting household object for observing how a policy behaves when its inputs differ from the demonstrations.
Conditions tried
These are qualitative observations from a small number of informal trials, not measured success rates.
| Variation | Condition | Observed behavior |
|---|---|---|
| Baseline | Trained book, fully visible | Completed pickup |
| Book identity | Unseen book, different cover and proportions | Completed pickup |
| Visibility | Book partially occluded | Completed pickup |
The policy also completed the partially occluded trial shown here. One possible explanation is that it relied on features beyond the full cover appearance, but these trials are too limited to determine which visual features it used.
ACT behaved differently in the desk cleanup experiments, where its behavior was more sensitive to changes in pen color and orientation. These are observations from two small, different datasets rather than a controlled comparison of the tasks.