PYTHON · NUMPY

Plan by imagining

Use the world model to choose actions — search plans in imagination, commit only the best to the real world.

01Challenge

Plan a torque sequence with the Cross-Entropy Method entirely inside the world model, then execute it on the real pendulum and reach the goal.

Predict firstA world model lets an agent plan by imagining rollouts. How many REAL environment steps does that planning cost?

02Model

Here is why world models are worth building: you can plan in them. Instead of trying actions on the real robot — slow, or dangerous, or impossible to undo — you imagine thousands of plans with the model, score them, and only commit the best one. This lesson runs the Cross-Entropy Method: sample action sequences, roll each out with the learned model, keep the elite few, refit, repeat. The whole search happens in imagination. Then you take the winning plan to the real pendulum and watch it arrive at the target angle it never physically explored. This is model-predictive control, and it is how world models turn prediction into action.

03Guided practice

1. A model good enough to plan with. A nonlinear world model over a torque range that can reach the goal. 2. CEM. Sample plans, score them by imagining, keep the elite, refit — search purely in the model. 3. Commit. Run the planned actions on the REAL pendulum; it reaches the target. Planning in imagination, acting once: model-predictive control.

04Feedback

PASS WHEN PASS when the imagined plan, executed for real, reaches the target far better than doing nothing.

If your run fails, check:

A cell errored — run them in order.
The planned actions should drive the real pendulum near the target.

05Retrieve & space

Why plan in a world model instead of on the real robot? (imagined trials are fast, safe, and undoable).
What does CEM keep between iterations? (the elite plans, to refit the sampling distribution).
What is this control scheme called? (model-predictive control — plan in the model, commit to the world).

06Mastery & project

You can turn a world model into a controller by searching for good plans in imagination.

When acting is expensive or risky, ask whether you can plan in a learned model first.

← Imagine in a latent A playable world →

PYTHON · NUMPY

Plan by imagining

Plan with CEM inside the model, then reach the goal on the real system.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you