Unit 13 · lesson
Autonomy Is Structured Decision-Making
Autonomy does not begin when a robot uses artificial intelligence.
A thermostat can make an autonomous decision. A line follower can behave autonomously. A warehouse robot can execute a mission without a person steering every motion.
The useful question is not "Is this AI?"
It is:
Which decisions can the robot make without a human issuing the next command?
A spectrum, not a switch
Consider four systems:
| System | Human role | Robot role |
|---|---|---|
| teleoperated rover | human chooses every motion | executes motor commands |
| assisted drive | human chooses goal/direction | robot limits or stabilizes motion |
| scripted autonomous | human starts mission | robot executes predefined states |
| adaptive autonomous | human defines mission/boundaries | robot chooses actions from current evidence |
The categories can overlap. The point is to locate decision authority.
Goal, state, action
A useful autonomy model is:
goal
↓
current state estimate
↓
decision rule / planner
↓
action
↓
new evidence
└──────────── back to state estimate
That is the same sense-decide-act loop from Unit 1, now with richer state and choices.
Autonomy needs boundaries
Suppose a delivery robot is told to reach a destination.
The mission does not automatically authorize it to:
- enter every area;
- exceed a speed limit;
- push through obstacles;
- continue with failed sensors;
- ignore a stop request.
Autonomy is always bounded by requirements.
Name the decisions
Choose one autonomous robot scenario.
List at least five decisions and mark each as:
- human-only;
- robot-allowed;
- robot-allowed only under a condition.
Then identify the evidence needed before the robot makes one of those decisions.
That list becomes the beginning of an autonomy contract.
Autonomy lives in a loop
An autonomous robot repeatedly answers versions of four questions:
Where am I?
What is happening?
What should happen next?
What command should I issue?
Those questions may be implemented by several subsystems, but the behavior emerges from the loop.
For a hallway delivery robot:
localize → check route → check obstacle → choose action → move → repeat
If the obstacle check becomes stale, the mission logic can make a perfectly logical decision from invalid evidence. Autonomy therefore depends on data validity, not only decision logic.
Separate mission goals from control actions
"Deliver the package to Room 204" is a mission goal.
"Command left wheel to 0.4 m/s" is a low-level control action.
Between them are intermediate decisions:
mission goal
↓
route / behavior
↓
motion target
↓
controller
↓
actuator command
Keeping those layers separate makes it easier to stop, recover, or change plans without rewriting motor-control code.
For an autonomous behavior you know, identify one decision at each level. If all decisions are collapsed into one giant loop, the architecture is probably hiding important responsibilities.
process flow
Automation and Autonomous Missions: Engineering Evidence Flow
Plan
Name the system, criterion, constraint, and safety condition.
Model
Trace the control, energy, and feedback paths.
Test
Run a bounded approved test and record evidence.
Revise
Document correction, limitation, and next safe action.
Read this concept flow as plain text
- Plan. Name the system, criterion, constraint, and safety condition.
- Model. Trace the control, energy, and feedback paths.
- Test. Run a bounded approved test and record evidence.
- Revise. Document correction, limitation, and next safe action.