Intelligent-Control Architectures & Timescale Hierarchies

An intelligent controller must do more than compute a feedback law: it may interpret sensors, maintain a model of the world, choose goals, plan actions, learn from outcomes, and explain decisions. Putting all of that into one undifferentiated loop makes timing, authority, and failure behavior impossible to reason about. A sound architecture therefore keeps fast physical regulation local and gives slower reasoning layers explicit, bounded ways to influence it.

Intelligence belongs in a control hierarchy #

A useful decomposition follows the spatial and temporal scale of decisions. At the bottom, servo or regulatory loops react in milliseconds to measured error. Above them, coordination logic changes modes, trajectories, or setpoints over longer horizons. Planning layers reason about tasks, resources, and alternatives over still longer horizons.

flowchart TD
  G[Mission and operator intent] --> P[Task planning and value judgment]
  P --> S[Supervisory modes and trajectories]
  S --> L[Local feedback controllers]
  L --> A[Actuators and physical process]
  A --> E[Fast sensors and state estimation]
  E --> L
  E --> W[World model and situation assessment]
  W --> S
  W --> P

The arrows are contracts, not suggestions. A planner may request a trajectory; it should not write actuator current directly. A local controller may reject or saturate a request that violates its envelope. The world model can aggregate observations for deliberation while the state estimator retains the bandwidth required by the fast loop.

NIST's Real-time Control System (RCS) expresses each level as a recurring combination of sensory processing, world modeling, value judgment, and behavior generation. The exact module names are less important than the invariant: perception, state, decisions, and actions remain traceable at every level.

Timescale is an engineering boundary #

Suppose a motor current loop runs at 20 kHz, a velocity loop at 1 kHz, a path follower at 100 Hz, and a task planner at 2 Hz. These rates differ because the controlled phenomena differ in bandwidth and because computation at each level has a different deadline.

The numbers are application-specific; the ordering is the point. A slower layer may set a target for a faster one, but it cannot assume that its own update rate is adequate for the plant dynamics.

Reactive and deliberative behavior solve different problems #

Reactive behavior maps a recognized situation to a bounded action with low latency. Deliberative behavior predicts alternatives and selects among them. Neither subsumes the other:

• pure reaction is fast but can be shortsighted or oscillate between locally attractive actions;
• pure deliberation can be too slow and depends on a world model that is always incomplete;
• a hybrid architecture lets reactive layers enforce immediate constraints while deliberative layers revise goals and plans.

This is why an autonomous robot still needs conventional feedback, interlocks, and state estimation. A learned planner or rule base does not replace motor control, collision monitoring, or actuator limits.

The world model is operational state, not a database dump #

A world model contains the state needed to select and evaluate behavior: estimated object poses, equipment modes, task progress, predicted consequences, uncertainty, and provenance. Its update policy must answer:

1. which observation changed the state;
2. when the observation was valid;
3. what uncertainty or confidence accompanies it;
4. which decisions consume it;
5. how contradictions and stale facts are exposed.

Without those contracts, a knowledge base can be internally consistent and still control yesterday's world.

Authority must narrow on the way down #

Higher layers usually have broader goals but less permission to manipulate hardware directly. Each boundary should define:

• accepted command types and units;
• feasible ranges and rate limits;
• deadlines, sequencing, and cancellation semantics;
• status and reason codes;
• which lower layer owns safety-related enforcement;
• what evidence is recorded for later diagnosis.

Worked decomposition: autonomous material handling #

Consider a mobile manipulator moving a tray between work cells.

• The mission layer chooses the next delivery from production priorities.
• The task layer decomposes it into navigate, align, grasp, transport, and release.
• The motion layer produces collision-checked paths and arm trajectories.
• Local loops regulate wheel torque, joint current, and gripper force.
• Perception updates obstacle and workpiece state; safety-rated sensing independently enforces protected-space constraints.

A blocked aisle should cause the motion layer to stop requesting forward velocity and the task layer to replan. It should not require the current controller to understand warehouse topology. Conversely, a current limit should be enforced locally even if the planner's model predicts the motion is safe.

Architecture review questions #

• Can every command be traced to a goal and every action to a bounded authority?
• Does each loop have an explicit update rate, deadline, and stale-data rule?
• Are estimation uncertainty and model age visible to decision logic?
• Can a slow or failed reasoning component block a fast safety-relevant loop?
• Are mode transitions atomic, observable, and testable?
• Do simulations include communication delay, dropped updates, saturation, and contradictory observations?

The value of an intelligent-control architecture is not the number of AI techniques it can host. It is the discipline with which it makes perception, reasoning, learning, and physical action composable and auditable.