Hierarchical Mind Design: Separating Cognition from Navigation
Keywords: Hierarchical Reinforcement Learning, Cognitive Architecture, Active Inference, World Model, Artificial General Intelligence, AGI, Cognitive Science, Navigation, Project Halo
When you decide to make a cup of coffee, your conscious mind does not plan the individual contractions of your leg muscles. It does not calculate the angle of your wrist as you open the cupboard. Your executive mind decides on the high-level intent ("make coffee"), and delegates the execution to lower-level sensorimotor loops that handle the physics of walking, reaching, and grasping.
If you try to plan everything at the same level—mixing high-level logic with low-level muscle control—your brain quickly deadlocks under the weight of the calculations.
In Volume 3, Episode 8, we solved context-dependent reasoning using version-space planning. But when we folded that cognitive core into the navigating gridworld agent (--gate15), we hit a performance wall.
The agent was starving, despite having a perfect understanding of the recipes.
We analyzed the traces and found that the bottleneck was navigation economy. The agent's sensory view was limited to adjacent tiles (a "smell" reflex). If it was far from a resource, the concentration gradient was flat, and it had no hill-climbing signal. It would wander aimlessly for 70 steps, burning its entire metabolic budget before it could assemble a recipe.
We needed a hierarchical architecture that separated high-level cognitive decisions from low-level spatial navigation.
The Hierarchical Split
We refactored the mind into two distinct tiers:
- The Cognitive Core (Version Space): Verbatim from Gate 14. It operates on abstract states, keeping track of which recipe hypotheses are active. It decides what the target action should be (e.g., "we need to probe Resource 1").
- The Body Navigator: A spatial affordance engine. It takes the cognitive core's intent as a command and handles the physics of pathfinding and gathering.
By separating the two, we created a firewall: rule-inference and navigation never touch.
The version space does not need to know where resources are on the grid; it only needs to know what resources exist. The navigator does not need to know why we are gathering Resource 1; it only needs to know how to get there safely.
[Cognitive Core] ──> Decides Intent (e.g., "Probe Resource 1")
│
└──> [Body Navigator] ──> Executes Pathfinding & Gathering
Debugging the Integration
Even with the hierarchical split, the first integration run failed. By measuring the system variables, we caught and resolved three distinct bugs:
- Cost Modeling Alignment: The cognitive core was planning based on abstract cost assumptions, causing it to command "single-parades" (trekking back and forth across the map for cheap probes, which actually cost massive metabolic energy). We aligned the planner's cost model with the navigator's real-path distance.
- Evidence-Gated Intent Stickiness: At every step, the agent's sensors received new observations, triggering the belief updater and causing it to flip-flop its intent (thrashing between different resources). We implemented a stickiness rule: the agent holds its active intent until the target resource is gathered or the hypothesis is falsified.
- The Dominance Rule: The navigator was absorbing resources whenever the local concentration cleared a threshold, which sometimes grabbed the wrong resource near overlapping sources. We corrected the navigator to only absorb the target when it was the locally dominant resource.
The Gated Results: Emergent Competence
Once these integration fixes were implemented, the survival rates climbed:
| Configuration | Survival Lifespan (Cap 600) |
|---|---|
| Old Broad Agent (v3 ref) | $100$ - $309$ steps |
| Hierarchical VS Agent | $317$ - $517$ steps |
The hierarchical mind successfully resolved the multi-recipe curriculum with a single set of weights, beating the old model-free agent by over $200$ steps of survival.
The Next Step: Giving the Mind Eyes
The hierarchical split was a major victory. It proved that separating cognition from navigation protects the agent from metabolic starvation.
However, the navigator was still relying on local gradient climbing. It had no spatial memory of where sources were, and no way to plan optimal paths around hazards. It was still walking blind.
To close the navigation gap once and for all, we needed to give the agent eyes. We needed a spatial mapping and path-planning organ.
In Volume 4, Episode 10, we look at how we built a persistent allocentric map and route planner under a frozen decision core, closing the navigation-economy wall.