World Models and Learned Simulators Quiz

Answers and explanations appear as you take the quiz.

1. What is the core job of a world model?
   • To predict how the environment changes given a state and action
   • To label images with categories
   • To compress text into embeddings
   • To route network packets
2. In the 2018 World Models DoomTakeCover experiment, where was the controller trained?
   • Only on real hardware
   • Inside the model's own learned 'dream' of the environment
   • On a public cloud cluster of robots
   • Using only language data
3. What notable feat did DeepMind's DreamerV3 achieve?
   • Writing a novel
   • Translating 100 languages
   • Collecting diamonds in Minecraft using policy learning from imagined trajectories
   • Designing a microchip
4. What does Google's Genie do?
   • Compresses videos for streaming
   • Sorts search results
   • Trains language tutors
   • Generates controllable, playable 2D worlds from images and unlabeled video
5. How do many world models keep the prediction problem tractable for high-dimensional inputs like images?
   • They encode observations into a compact latent state and predict in that space
   • They ignore the images entirely
   • They use only audio
   • They store every raw frame forever
6. What is a potential motivation for using a world model in robotics?
   • Guarantee every simulated action is safe to deploy
   • Explore candidate actions in learned dynamics with fewer real-world interactions, while validating model accuracy
   • Avoid testing the model against any real observations
   • Prove a visually realistic frame is physically accurate
7. What techniques did DreamerV3 report for improving learning stability across diverse domains?
   • A fixed image resize and no reward modeling
   • A hand-written controller for every game
   • Normalization, balancing, and transformations
   • Removing all state observations from training
8. Why does visually plausible output from a learned simulator not by itself prove accurate dynamics?
   • Visual models cannot process actions
   • Every rendered image is a ground-truth sensor measurement
   • Plausibility guarantees accurate forces and collisions
   • A realistic-looking frame may still depict incorrect physical state changes