Module 4/5 · Weeks 10–12 · 27 h

Intelligent control

DRT 344 Automation, Robotics and Intelligent Control Systems

About 90 minDraft, awaiting reviewLast updated 26 September 2026

By the end of this module you will be able to

Explain the perceive, understand, plan and control loop of autonomous systems
Explain levels of automation and the role of human oversight
Compare the Dijkstra, A* and RRT path-planning algorithms
Calculate the output of a simple fuzzy controller
Explain the principles of MPC and reinforcement learning and their safety limits

Prerequisites: DRT 344 modules 2–3

Why this matters

The PID controller of module 3 keeps a drone where it is told to be, but who decides where it should go? What if a tree is in the way, the target moves, or the data are uncertain? Advanced automated systems must perceive their environment, plan and decide for themselves. Together this is called intelligent control and autonomy.

Think of a driver. PID is the skill of steering within the lane; intelligent control is reading signs, choosing a route and deciding to brake when someone steps into the road.

The autonomy loop

A four-step loop: perceive with sensors and AI, understand with a map and localisation, plan the path and behaviour, control and act, then back to perceive; human oversight sits in the centre — Figure 1 The autonomy loop

Perceive: read sensors and use AI to detect objects
Understand: build a map and localise
Plan: choose a path and behaviour
Control: command low-level controllers such as PID to follow the plan

Every step carries uncertainty, and a perception error propagates through every later step. A good system therefore has human oversight at a level appropriate to the risk.

Levels of automation

Autonomy is not just “piloted” versus “flies itself”; there are many levels. Useful reference frameworks include:

NIST ALFUS (Autonomy Levels for Unmanned Systems, SP 1011), which characterises unmanned-system autonomy by mission complexity, environmental difficulty and the degree of reliance on humans
SAE J3016, which divides road-vehicle driving automation into levels 0–5, from no automation to full automation. It is a helpful analogy, though not designed for drones

The key questions at each level are who is accountable when the system decides wrongly, and does the human know in time when to intervene (linking to ethics in DRT 343 module 5).

Path planning

A grid map divides space into cells; obstacle cells cannot be entered. Search algorithms find a path from start S to goal G.

Dijkstra (1959) expands nodes in order of accumulated cost from the start, $g (n)$ . It guarantees the shortest path but searches in every direction
A* (Hart, Nilsson & Raphael, 1968) uses $f (n) = g (n) + h (n)$ , where $h (n)$ estimates the remaining distance to the goal. If $h$ never overestimates, A* still finds the shortest path while expanding far fewer nodes
RRT (LaValle, 1998) randomly grows a tree of paths through the space. It suits continuous, high-dimensional spaces such as a drone in 3D, but its paths are usually not shortest and need smoothing afterwards

An eight by eight grid with a vertical wall of four obstacle cells; start S on the left and goal G on the right; yellow cells were expanded by A*, green cells were not; a pink line shows the 11-step shortest path going around below the wall — Figure 2 A* path search on a grid

Example 1 A* versus Dijkstra

An 8×8 grid, 4-connected, 1 unit per step, with a vertical wall of 4 cells between S and G, using Manhattan distance as the heuristic.

At the start $h (S) = ∣4 - 4∣ + ∣0 - 7∣ = 7$ , but the wall blocks the straight route, so the real path must detour
Results from a Python program: both methods find a path of 11 steps
Dijkstra expands 59 nodes; A* expands only 18

A good heuristic greatly reduces computation, which matters on a drone’s limited onboard computer that must replan whenever it finds a new obstacle.

Fuzzy control

Zadeh’s (1965) fuzzy logic gives set membership values between 0 and 1 instead of yes or no. Mamdani and Assilian (1975) used it to build controllers from human-language rules such as “if far from the target, fly fast”.

Example 2 A simple fuzzy speed controller

Rules: if the distance is near, fly at 2 m/s; if far, fly at 8 m/s. Over 0–10 m, let $μ_{near} = (10 - e) /10$ and $μ_{far} = e /10$ . The drone is now $e = 6$ m from the target.

$μ_{near} = (10 - 6) /10 = 0.4$ and $μ_{far} = 6/10 = 0.6$
$u = \frac{0.4 \times 2 + 0.6 \times 8}{0.4 + 0.6} = 5.6$ m/s

The speed changes smoothly with distance instead of jumping from 2 to 8 m/s at a single boundary.

MPC and reinforcement learning

Model predictive control (MPC) uses a model to predict the system over a horizon and finds the best commands subject to constraints, such as maximum speed or area boundaries. It applies the first command and recomputes every cycle. Its strength is handling constraints directly; its cost is heavy computation
Reinforcement learning (RL) lets an agent learn a policy by trial and error to maximise reward (Sutton & Barto, 2018). A landmark example: Kaufmann et al. (2023) trained a racing drone with deep RL that beat human champions on a test course, training mostly in simulation

Learning systems carry a particular risk: they work well in conditions they were trained on but are hard to predict in unfamiliar ones. Real deployment needs verifiable protection layers, such as geofences and non-AI failsafes (DRT 342 module 2), SITL testing first, and human oversight.

Class activity

Activity: Planning a path and testing in simulation

Implement A* on a grid in Python, move the obstacles, and compare the nodes expanded with Dijkstra.
Try an overestimating heuristic such as $2 h$ and check whether the path is still shortest.
Design fuzzy rules to slow down when approaching an obstacle.
Discuss which protection layers an AI avoidance system needs before real flight, drawing on the SITL knowledge unit in the drone knowledge hub.

Common mistakes

Watch out

Using an overestimating heuristic and assuming the path is still shortest
Planning on an old map without replanning when obstacles appear
Trusting AI output without verifiable protection layers
Misreading levels of automation as meaning an autonomous system needs no accountable person
Deploying an RL policy trained in simulation directly without testing the sim-to-real gap

Summary

Autonomous systems run a loop of perceive → understand → plan → control, with human oversight
Autonomy has many levels; NIST ALFUS and SAE J3016 are reference frameworks
A* uses $f = g + h$ , finds the same shortest path as Dijkstra while expanding fewer nodes; RRT suits continuous high-dimensional spaces
Fuzzy control uses human-language rules with membership 0–1; MPC handles constraints directly
RL can achieve remarkable results but needs verifiable protection before real use

Check your understanding

In A*, what do $g (n)$ and $h (n)$ mean?
What is the Manhattan distance from $(1, 2)$ to the goal $(5, 7)$ on a grid?
In the fuzzy example, what speed is commanded at 3 m?
What is the main advantage of MPC over PID?
Why does an AI avoidance system still need a geofence and non-AI failsafes?

Answers

$g (n)$ is the actual cost from the start to node $n$ ; $h (n)$ estimates the remaining cost from $n$ to the goal
$∣1 - 5∣ + ∣2 - 7∣ = 4 + 5 = 9$
$μ_{near} = 0.7$ , $μ_{far} = 0.3$ , so $u = 0.7 \times 2 + 0.3 \times 8 = 3.8$ m/s
MPC predicts the future with a model and handles constraints directly
AI may fail in conditions it was not trained on, so verifiable, predictable protection layers are required

Key formulas

A* evaluation function	$f (n) = g (n) + h (n)$
Manhattan distance (4-connected grid heuristic)	$h (n) = ∣ x_{n} - x_{G} ∣ + ∣ y_{n} - y_{G} ∣$
Weighted average (simple defuzzification)	$u = \frac{\sum _{i} μ _{i} u _{i}}{\sum _{i} μ _{i}}$

Key references

Huang, H.-M. (Ed.). (2008). Autonomy levels for unmanned systems (ALFUS) framework, Volume I: Terminology, Version 2.0 (NIST SP 1011-I-2.0). NIST. link
SAE International. (2021). Taxonomy and definitions for terms related to driving automation systems for on-road motor vehicles (J3016_202104). link
Dijkstra, E. W. (1959). A note on two problems in connexion with graphs. Numerische Mathematik, 1, 269–271. link
Hart, P. E., Nilsson, N. J., & Raphael, B. (1968). A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions on Systems Science and Cybernetics, 4(2), 100–107. link
LaValle, S. M. (1998). Rapidly-exploring random trees: A new tool for path planning (TR 98-11). Iowa State University. link
Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338–353. link
Mamdani, E. H., & Assilian, S. (1975). An experiment in linguistic synthesis with a fuzzy logic controller. International Journal of Man-Machine Studies, 7(1), 1–13. link
Rawlings, J. B., Mayne, D. Q., & Diehl, M. M. (2017). Model predictive control: Theory, computation, and design (2nd ed.). Nob Hill Publishing. link
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction (2nd ed.). MIT Press. link
Kaufmann, E., Bauersfeld, L., Loquercio, A., Müller, M., Koltun, V., & Scaramuzza, D. (2023). Champion-level drone racing using deep reinforcement learning. Nature, 620, 982–987. link
Siegwart, R., Nourbakhsh, I. R., & Scaramuzza, D. (2011). Introduction to autonomous mobile robots (2nd ed.). MIT Press.
Thrun, S., Burgard, W., & Fox, D. (2005). Probabilistic robotics. MIT Press.

In class / field

Lecture, case discussion and in-class problem solving

Learning evidence: Quiz results and submitted exercises

Module quiz

This is a formative self-check, not a graded exam

Knowledge domain: Control, autopilot and navigation · Automation, robotics and swarms · Artificial intelligence and computer vision · Programming and digital technology · Surveying, mapping and geoinformatics · Mission planning, flight and simulation

Intelligent control

By the end of this module you will be able to

Why this matters

The autonomy loop

Levels of automation

Path planning

Fuzzy control

MPC and reinforcement learning

Class activity

Common mistakes

Summary

Check your understanding

Key formulas

Key references

Further reading

Autonomous path planning and obstacle avoidance

Topic 2: AI and automated control

Developing simulations and connecting SITL

In class / field

Module quiz

Intelligent control

Lesson

By the end of this module you will be able to

Why this matters

The autonomy loop

Levels of automation

Path planning

Fuzzy control

MPC and reinforcement learning

Class activity

Common mistakes

Summary

Check your understanding

Key formulas

Key references

Further reading

Autonomous path planning and obstacle avoidance

Topic 2: AI and automated control

Developing simulations and connecting SITL

In class / field

Module quiz