An object disappears behind an obstacle. Remembering it requires both storing its state and predicting where it goes while unseen. A large memory can fail if its motion assumptions are too narrow.
An October 13 comparison probes FloWM's spatial grid against an oracle-fed Kalman tracker in the model's block world. Results depend on how the grid is read. The tracker also handles diagonal motion that lies outside the grid's built-in velocity choices much better. Read the research paper.
The small baseline receives privileged information
That advantage needs careful interpretation. An oracle-fed tracker is not a ready replacement for a learned perception system. It isolates representation and motion structure by removing part of the sensing problem.
The comparison is valuable precisely because it separates these questions. A failure attributed to weak memory may originate in an inadequate transition model. Increasing storage would preserve the same mistaken forecast more expensively.
For robot world models, vary motion direction and speed while holding visibility gaps constant. Then vary sensing noise separately. These interventions identify whether the bottleneck is observation, state storage or dynamics.
The paper does not establish that hand-built trackers generally outperform learned world models. It shows why a well-chosen privileged baseline can reveal what an impressive memory score is actually measuring.
References
Questions, news tips or corrections? Contact [email protected].
