Researchers led by Ho Tim Tsui have studied how a drone's learned policy for landing on a moving platform copes with a landing pad that looks different from the one it was trained on, according to a paper on arXiv. The policy sees only binary masks of the target from a SAM 3 detector, which strip out most of the pad's appearance and let a text prompt say what to look for.
In simulation, a banana-themed pad cut landing success from 75.6% to 12.0%, because the detector rarely found it under the original prompt. A revised prompt, picked after testing several, restored success to 68.4% without changing the policy's weights.
In a separate test, filling in masks that were missing while the pad was in view helped landing far more than correcting the masks' position or size. Tracking the pad over time supplied masks more often but did not reliably improve landing, so whether a usable mask reaches the policy at all appears to be a key limit, the authors write.



