After studying this material, you should be able to:

1. Define topology and explain why it differs from standard geometric/spatial reasoning
2. Identify the five core topological properties tested in MindTopo
3. Distinguish between topological reasoning (static) and topological planning (interactive)
4. Explain why current AI models fail at topological tasks and where those failures occur
5. Connect these findings to real-world applications in robotics and AI systems

---

Step-by-Step Teaching

Step 1: What Is Topology? (Building the Foundation)

Before understanding MindTopo, you need to understand what topology actually means.

Topology is NOT about:

• Exact distances ("the chair is 2 meters away")
• Precise angles ("turn 45 degrees")
• Specific sizes or shapes

Topology IS about:

• Structural relationships that survive deformation
• Think of it this way: if you stretch, bend, or squish an object without cutting or gluing, topological properties remain unchanged

Concrete Analogy

Imagine a rubber band:

Original circle:    Stretched oval:    Twisted figure-8:
     O                  (   )               ∞
  
Still one loop     Still one loop     Now TWO loops — topology changed!

The number of loops is a topological property. Stretching doesn't change it. Cutting does.

Why This Matters for AI

Humans naturally track these structural relationships. You instantly know:

• Whether a door connects two rooms
• Whether a dog is inside or outside a fence
• Whether headphone cables are truly knotted or just tangled

AI systems have largely been tested on Euclidean properties (distance, direction, size) — not on these structural relationships.

---

Step 2: The Five Topological Properties in MindTopo

MindTopo organizes everything around five categories. Learn each one with a clear example:

Property 1: Continuity

Question it asks: Is this path or object unbroken?

Connected path:          Broken path:
A ——————— B              A ———   ——— B
                                ↑
                           gap here

Real example: Are two rooms still connected after a wall is added?

---

Property 2: Separation

Question it asks: Do these elements form one structure or distinct parts?

One structure:     Two separate parts:
  ∪∩               ∪    ∩
(linked)          (not linked)

Real example: Are these two rope loops linked together or completely separate?

---

Property 3: Order

Question it asks: How are elements arranged along a path or through a transformation?

Beads on a string:
A — B — C — D

Order matters: A,B,C,D ≠ A,C,B,D

Real example: After folding origami, which point ends up on top?

---

Property 4: Enclosure

Question it asks: Does a boundary create a clear inside and outside?

Sheep INSIDE fence:    Sheep OUTSIDE fence:
  ___                      ___
 |🐑 |                    |   |    🐑
 |___|                    |___|

Real example: Is the animal enclosed within the fence boundary?

---

Property 5: Knots

Question it asks: Is this rope truly knotted, or just tangled in appearance?

True knot:              Unknot (just tangled-looking):
  ___                        ___
 /   \                      /   \
|  ×  |   ← crossing       |     |   ← no real crossing
 \___/                      \___/
Cannot be undone            Can be undone without cutting
without cutting

Real example: Can you untangle these headphones without cutting them?

---

Step 3: How MindTopo Tests These Properties — Two Cognitive Levels

This is a critical distinction. MindTopo tests each property at two levels:

Level 1: Reasoning (Static)

• Model looks at a scene (image)
• Answers a question about its topological structure
• No action required — just perception and understanding

[Image of maze] → "Are points A and B connected?" → "Yes/No"
[Image of sheep] → "Is the sheep inside the fence?" → "Yes/No"
[Image of rope] → "Is this truly knotted?" → "Yes/No"

Level 2: Planning (Interactive)

• Model interacts with a simulated environment
• Must take a sequence of actions to create, preserve, or remove a topological relationship
• The environment enforces physical rules (you cannot pass a rope through itself)

[Pipe puzzle] → Rotate segments → Connect path from A to B
[Rope puzzle] → Move strands → Untangle without passing through each other
[Chat Noir]   → Place barriers → Trap the moving agent inside

Key Insight: These Are Very Different Skills

---

Step 4: What the Research Found — The Performance Gap

Finding 1: Static > Interactive (by a large margin)

AI models performed significantly better on reasoning tasks than planning tasks. Both remained well below human performance.

Performance (conceptual):