Can you solve this NVIDIA interview question?

Imagine a network that receives a drone’s recent states and actions and compresses them into a hidden vector. Should it reconstruct the current sensors, predict the next state, distinguish flights, or simply make its latent as small as possible? The course hypothesis is B: predicting consequences gives the representation a reason to retain dynamics. The interview framing comes from the course specification; no independent claim about an actual NVIDIA hiring question is established here.

A snapshot can say where a drone is. The response to a known action can say something about what the drone is. A heavier vehicle and a weaker motor may respond similarly, so prediction does not guarantee unique physical parameters. Informative maneuvers, careful data splits and explicit assumptions still matter.

The nine-module experiment builds the test rather than assuming its outcome. First establish physics and classical adaptation. Then train all four objectives, compare held-out probes and tracking, and bring the surviving models to a printed aircraft. These are written episode companions and experiment guides; there are no claimed flight results or video releases on this page.