Sussman Anomaly

Duration: 8 min

This video lesson is available to enrolled students.

Enroll to watch — ZERO TO HERO

AI Summary

An AI-generated summary of this video lecture.

This lecture covers the representation of goals and actions in STRIPS, using the Blocks World problem as a primary example. It introduces the Closed World Assumption and details the Sussman Anomaly, a classic planning problem where noninterleaved planners fail to achieve stacked goals. The session concludes by introducing Plan Space Planning (Non-Linear Planning) as a solution and begins defining actions using PDDL notation.

Chapters

  1. 0:00 – 2:00 00:00-02:00

    The lecture begins by defining goals in STRIPS as a set of facts. Using the Blocks World, it illustrates that `on(A, B)` is a goal state if present in the current state. State 1 is not a goal state because A is held, whereas State 2 is a goal state because A is on B. The instructor then transitions to 'Representing Actions', explaining that actions are defined by a set of preconditions, add effects, and delete effects. Handwritten notes label 'Robot', 'Action', and 'Hand' to contextualize the components.

  2. 2:00 – 5:00 02:00-05:00

    The instructor details the Blocks World representation, stating that states are sets of facts or propositions. For State 1, the facts listed are `on(C, Table)`, `on(B, C)`, `clear(B)`, and `holding(A)`. The concept of 'Closed World Assumption (CWA)' is introduced, noting that facts not listed are assumed false. The lecture then introduces the 'Sussman Anomaly' slide, showing an initial state with C on A and B on the table, and a final state with A on B on C.

  3. 5:00 – 7:47 05:00-07:47

    The Sussman Anomaly is explained: noninterleaved planners separate goals like 'A atop B' and 'B atop C'. Achieving one goal often undoes the other, creating a conflict. The solution proposed is 'Plan Space Planning' or 'Non-Linear Planning', which allows modifying any part of the plan and shifting attention midway. Finally, the instructor introduces PDDL, defining the `Move(b, x, y)` action with preconditions like `Clear(b)` and effects like `On(b, y)`.

The video progresses from fundamental STRIPS concepts to complex planning challenges. It starts by establishing how goals and actions are formally represented using sets of facts and logical conditions. This foundation leads to the Blocks World problem, where the instructor highlights the Closed World Assumption. The core difficulty, the Sussman Anomaly, is then presented as a failure of linear planners to handle interdependent subgoals. The lecture resolves this by introducing Plan Space Planning, which allows for non-linear execution, and begins the technical specification of actions using PDDL syntax.