AI Notes
Learn autonomous systems. Decision making, learning. Robotics 2024.
What Are Autonomous Systems?
An autonomous system is a machine that can sense its environment, make decisions, and take actions to achieve goals without continuous human intervention. The degree of autonomy ranges from simple automation (thermostat) to full autonomy (self-driving vehicle navigating novel situations). True autonomy requires integrating perception, reasoning, planning, and execution in a closed loop that handles uncertainty and novel situations gracefully.
The autonomy spectrum is commonly defined using SAE levels for vehicles, but applies broadly: Level 0 (no autonomy—human does everything), Level 2 (partial automation—system assists), Level 4 (high automation—system handles most situations), Level 5 (full autonomy—no human needed ever).
Architecture of Autonomous Systems
Self-Driving Vehicles: Case Study
Perception Stack
Sensor Suite
- LiDAR (360° 3D point cloud, 10Hz, 100m range)
- Cameras (8+ cameras, 360° coverage, traffic lights/signs)
- Radar (long-range, works in rain/fog)
- Ultrasonic (close-range parking)
- GPS/IMU (global position + motion)
Sensor Fusion
Combine all sensors into unified world model
Each sensor compensates for others' weaknesses
LiDAR fails in heavy rain → radar fills in
Cameras can't measure distance → LiDAR provides depth
GPS loses signal in tunnel → IMU dead reckons
Detection Pipeline
Point cloud + images → 3D object detection
Outputs: [car at (x,y,z), pedestrian at (x,y,z), ...]
Each with velocity estimate and uncertainty bounds
Planning & Decision Making
| Behavior Planning | What to do |
| Motion Planning | How to do it |
| Prediction | What will others do |
Autonomous Drones (UAVs)
Applications
- Package delivery (Amazon Prime Air)
- Agriculture (crop spraying, monitoring)
- Search and rescue (finding lost hikers)
- Infrastructure inspection (power lines, bridges)
Key capabilities
- Visual navigation (no GPS indoors/underground)
- Obstacle avoidance (trees, wires, birds)
- Path planning with energy constraints (battery life)
- Multi-drone coordination (swarm intelligence)
Autonomy challenges specific to drones
- 3D planning (altitude adds dimension)
- Wind compensation (constant disturbances)
- Limited compute (lightweight hardware)
- Regulatory compliance (no-fly zones, altitude limits)
Decision-Making Under Uncertainty
| Challenge | Sensor noise + unpredictable environment |
| Approach | Probabilistic planning |
| Example | Should the car merge? |
| Merge now | 0.85×(+10 saves time) + 0.15×(-50 danger) = 1.0 |
| Wait | 0×risk + 1.0×(+5 slightly late) = 5.0 |
| Decision | Wait (higher expected utility considering safety) |
| Key principle | Autonomous systems must be conservative |
Safety and Verification
Formal Verification
Mathematically prove safety properties hold
"Robot never enters unsafe state" — proved for all inputs
Runtime Monitoring
Independent safety system monitors behavior
Can override autonomous decisions if safety violated
"Watchdog" that triggers emergency stop
Redundancy
Multiple independent systems must agree
Triple-redundant flight computers (aircraft)
If 2/3 agree and 1 disagrees → follow majority, flag anomaly
Testing
Billions of simulated miles
Thousands of real-world test miles
Formal methods for critical subsystems
Edge case generation (adversarial testing)
Ethical Considerations
Trolley Problem in Autonomous Vehicles
Unavoidable accident: hit pedestrian or swerve into wall?
Current approach: minimize harm probability, not choose victims
Legal framework: still evolving worldwide
Accountability
When an autonomous system causes harm, who is responsible?
Manufacturer? Software developer? Owner? The system itself?
Transparency
Can the system explain its decisions?
Required for trust, debugging, and legal compliance
Bias
Training data may underrepresent certain populations
Detection systems may work worse for some groups
Must actively test for and mitigate biases
Interview Questions
Q: What is the difference between autonomous and automated systems? A: Automated systems follow pre-programmed rules for known situations (assembly line robot repeating the same task). Autonomous systems can handle novel situations through perception and reasoning (self-driving car navigating construction zones it's never seen). Autonomy implies adaptability; automation implies repeatability.
Q: Why is Level 5 autonomy so difficult to achieve? A: The "long tail" of edge cases: unusual situations that rarely occur but require sophisticated reasoning (construction worker gesturing traffic, emergency vehicle approaching from unusual direction, road debris). Each edge case is rare, but collectively they occur frequently. No amount of training data covers every possible situation.
Q: How do autonomous systems handle sensor failure? A: Through redundancy and graceful degradation. If one sensor fails, others compensate (camera fails → rely on LiDAR + radar). If critical perception is impaired, the system enters a "safe stop" mode—pulling over, slowing down, or alerting a human operator. The architecture must be designed so no single sensor failure causes catastrophe.
Exam Focus
Revise definitions, diagrams, examples, and short-answer points for Autonomous Systems - Self-Directed Robots.
Interview Use
Prepare one clear explanation, one practical example, and one common mistake for this Artificial Intelligence topic.
Search Terms
artificial-intelligence, artificial intelligence, artificial, intelligence, robotics, autonomous, systems, autonomous systems - self-directed robots
Related Artificial Intelligence Topics