← Course schedule · WEEK 06

Generative policies: interactive demos

Explore multimodality, diffusion and flow matching, policy steering, and action-chunk execution. Choose a demo below, then use its walkthrough and controls to explore.

Lecture 06 — Generative Policies slides (PDF)

Open all 14 demos on one page →

1. What should the policy predict?

   Two valid routes, their conditional mean, and a sampled behavior.

2. What does adding noise do to a distribution?

   Track samples and density as the signal-to-noise ratio changes.

3. Can we turn noise into a useful sample?

   An exact toy denoiser and a deterministic DDIM sampler.

4. What does the network actually learn?

   Inspect real checkpoints of a small trained noise-prediction network.

5. How many model calls do we need?

   A fixed initial sample and one checkpoint, with different step budgets.

6. Denoising an entire action chunk

   A learned waypoint distribution decoded into a synthetic action path.

7. Same noise, different observations

   Hold the sampling seed fixed while changing the policy condition.

8. Learning a velocity field

   Straight training pairs and trajectories from a trained flow model.

9. Which parts of an imperfect demonstration are useful?

   Noise level changes the visibility of local motion errors.

10. One policy, several noise inputs

    Inspect different seeds, then reuse one across changing observations.

11. Choosing where generated actions should go

    An exact toy product distribution illustrates guidance and composition.

12. The robot keeps moving during inference

    Observe, compute, and execute on separate timelines.

13. Connecting the next chunk to the current one

    Committed actions, soft overlap, and a free future.

14. Smoothness and reactivity are different measurements

    Explore how overlap strength changes continuity and response.