Diffusion Policy for Robot Control Quiz

Answers and explanations appear as you take the quiz.

1. What does a Diffusion Policy generate at each decision point?
   • A short sequence (chunk) of future actions
   • A single instantaneous motor command
   • A natural-language explanation
   • A reward signal for training
2. Which generative technique does Diffusion Policy borrow from image AI?
   • Manual pixel painting
   • Iterative denoising of random noise
   • Genetic algorithms
   • Decision trees
3. How can a generative action policy represent a task with several valid solutions?
   • It must choose the same fixed action for every observation
   • It removes the need for demonstrations or action data
   • It can model multiple plausible action sequences instead of regressing to their average
   • It guarantees higher success than every older controller
4. During training, what is the network in a Diffusion Policy taught to predict?
   • The next word in a sentence
   • The robot's battery level
   • The camera's exposure setting
   • The noise that was added to a demonstrated action sequence
5. What is 'receding-horizon' replanning in Diffusion Policy?
   • Predicting an action chunk but re-planning with fresh observations before finishing it
   • Running the robot only once then stopping
   • Ignoring all sensor input
   • Training without demonstrations
6. Which architecture did the Diffusion Policy paper introduce for modeling temporal action sequences?
   • A static image classifier with no temporal model
   • A time-series diffusion transformer
   • A text-only recurrent language model
   • A fixed lookup table of robot motions
7. What observations can condition a visuomotor Diffusion Policy?
   • Only a text prompt, with no robot or scene data
   • The names of the training researchers
   • Visual observations and the robot’s state
   • A future action sequence labeled as correct at inference time
8. How broad was the Diffusion Policy paper’s benchmark evaluation?
   • One task on a single robot benchmark
   • Hundreds of tasks across every robot platform
   • Only simulations with no manipulation tasks
   • Twelve tasks across four robot-manipulation benchmarks