Can you solve this NVIDIA interview question?

Give a network the last half second of an arm’s servo readings and goals, and ask it to compress them into a hidden vector. Should the vector reconstruct the present readings (A), predict the next state from state and action (B), tell runs apart (C), or simply stay as small as possible (D)? The course answers B: prediction gives the vector a direct reason to keep what a payload, friction or a softer servo does to the next moves, while the other objectives reward that only indirectly, if at all. The interview framing comes from the course spec; this page establishes no claim about an actual NVIDIA hiring question.

A snapshot says where the arm is. How the arm answers a known goal says something about what it carries. On this arm the answer has a known limit: the motor currents reveal payload mass only relative to the servos’ torque constant, so a heavier cup and a weaker servo can look alike, and the test targets their ratio, m/kₜ. Module 4 trains one encoder with all four objectives on the same data and budget, probes each latent on held-out arms, and reports all four, whatever they show.

The nine modules build that test instead of assuming its outcome: a simulator first, then classical control and an EKF as baselines, then the four objectives and a policy that adapts, then a printed arm the model has never seen. That is the promise: to test B on a real robot arm, where a tuned classical controller may still win, and a classical win would count as a result. These are written episode companions and experiment guides; there are no claimed results or video releases on this page.