Physical AI Lab // Episode 3.2 // Quiz
World models Making data with Cosmos
Predicting the future (Predict) and transferring style (Transfer)
01
Which statement is closest to the video's definition of a "world model"?
Give it past frames and an action, and it predicts the next moment
An engine that renders frames from physics rules people have written
A video tokenizer that compresses video into a compact string of tokens
A planner that rehearses several actions and picks the best one
02
You have segmentation and depth maps rendered in simulation, and the frames look obviously fake. To turn them into realistic training frames, which tool do you use?
Cosmos Predict: give it a first frame and let it predict the frames that follow
Cosmos Transfer: conditioned on segmentation and depth, it makes real frames
The video tokenizer: compress the sim frames into tokens and decode them as real
The guardrails: block the unrealistic frames first and keep only the real-looking ones
03
Someone cites "Cosmos scored 78.3% on robot tasks" to argue it brings robot task success close to eighty percent. What's wrong with this?
The number is misremembered; the paper's figure here is actually 13.0%
78.3% is physical accuracy in simulation, not in real scenes
78.3% is human reviewers' overall preference, not robot task success
78.3% is a peak measured on a top-end rack; normal setups are far lower
04
After augmenting data with generated video, you check frame by frame and find a cup vanishes midway, then reappears out of nowhere. What does this tell you, and what do you do?
The video tokenizer compressed too hard; regenerate at a higher bitrate and it won't recur
A common world-model physics inconsistency: log it as a failure, not ground truth
The prompt wasn't detailed enough; a fuller description will guarantee correct physics
The downstream policy has already learned badly, so retrain the policy first
05
Which line best sums up how the video positions generating training data with a world model?
Once the frames look real enough, use them as ground truth and skip real-robot collection
It mainly serves self-driving and does little for robot manipulation
With enough compute for real-time generation, physical inconsistencies disappear
It's a data amplifier: cheap and varied, saving real-robot time but not the checking
Check