4.1 Levels of Automation

Autonomous flight systems range from pilot assistance to fully autonomous operations. The industry adapts automotive SAE levels (0-5) to aviation context, with most current eVTOLs targeting Level 3-4 initially before progressing to Level 5.

Level	Name	Description	Human Role	Example eVTOL
0	Manual	Pilot controls all functions	Full control	Early prototypes
1	Assisted	Autopilot holds altitude, heading	Constant monitoring	Volocopter (initial)
2	Partial	Automated flight segments	Ready to take over	Joby S4 (pilot-assist)
3	Conditional	Full automation in specific conditions	Available if needed	Archer Midnight
4	High	Full automation in defined routes	Passenger oversight	EHang 216
5	Full	Complete autonomy all conditions	None required	Wisk Cora (target)

4.2 Sensor Suite Architecture

Autonomous eVTOLs employ redundant sensor arrays providing 360-degree environmental awareness. Wisk's Generation 6 aircraft demonstrates state-of-the-art sensor fusion:

• GPS/GNSS (Triple-redundant): Primary navigation with RTK (Real-Time Kinematic) achieving centimeter-level accuracy
• Inertial Measurement Units (6x): Accelerometers and gyroscopes tracking motion in all axes, cross-validated for accuracy
• Radar (4x): 77 GHz automotive-grade radar detecting obstacles, terrain, and other aircraft in all weather
• LiDAR (3x): Laser scanning creating 3D point clouds for precision obstacle avoidance and terrain mapping
• Cameras (12x): Visual spectrum and infrared providing 360° coverage for computer vision algorithms
• Ultrasonic Sensors (8x): Short-range proximity detection during landing and ground operations
• Air Data Sensors (6x): Measuring airspeed, altitude, angle of attack, and sideslip
• Weather Sensors: Wind speed/direction, temperature, humidity, precipitation detection

Sensor Fusion

Individual sensors have limitations—GPS fails indoors, cameras struggle in fog, radar can't identify objects. Sensor fusion algorithms combine data from all sources, creating robust understanding even when individual sensors malfunction or encounter adverse conditions.

Kalman filters mathematically combine sensor measurements weighted by confidence levels. If GPS accuracy degrades near tall buildings, the system relies more heavily on visual-inertial odometry from cameras and IMUs. Machine learning models identify and dismiss sensor readings inconsistent with physics or other measurements.

4.3 Flight Control Algorithms

Autonomous flight controllers manage aircraft motion through sophisticated control loops operating at 100-1000 Hz. The control architecture typically follows a hierarchical structure:

High-Level Mission Planning

Route optimization algorithms calculate optimal flight paths considering weather, airspace restrictions, terrain, and energy efficiency. Dynamic replanning occurs continuously as conditions change. Machine learning models trained on millions of simulated flights predict optimal routes better than traditional algorithms.

Mid-Level Guidance

Guidance systems generate desired aircraft states (position, velocity, attitude) to follow the planned route. Model Predictive Control (MPC) anticipates future states multiple seconds ahead, optimizing control inputs for smooth, efficient flight. Constraints include passenger comfort (limiting acceleration), energy consumption, and safety margins.

Low-Level Stabilization

Fast inner loops stabilize aircraft attitude (roll, pitch, yaw) hundreds of times per second. PID (Proportional-Integral-Derivative) controllers adjusted by gain scheduling adapt to different flight regimes. During transitions, advanced nonlinear control maintains stability as aerodynamics change dramatically.

4.4 Detect and Avoid Systems

Technology	Detection Range	All-Weather	Processing Load	Limitations
Computer Vision	50-500 meters	No (weather dependent)	Very High	Requires clear visibility, compute-intensive
Radar	200-2000 meters	Yes	Moderate	Poor resolution, can't identify object types
LiDAR	100-300 meters	Partial (rain/fog limits)	High	Limited range, affected by precipitation
ADS-B	40+ kilometers	Yes	Low	Only equipped aircraft, no terrain/obstacles
Acoustic	50-200 meters	Yes	Low	Short range, directionality challenges

Effective collision avoidance requires multiple complementary technologies. ADS-B transponders detect other aircraft broadcasting their position. Radar provides all-weather capability. Computer vision identifies birds, drones, and non-cooperative aircraft. LiDAR maps terrain and obstacles with high precision.

Collision avoidance algorithms calculate closest point of approach (CPA) for all detected objects. If CPA violates safety margins (typically 500 feet horizontal, 250 feet vertical), the system executes avoidance maneuvers automatically. Priority rules determine right-of-way, following aviation conventions.

4.5 Machine Learning and AI

Modern eVTOLs leverage machine learning throughout autonomous systems:

Perception

Convolutional neural networks (CNNs) trained on millions of labeled images identify objects in camera feeds—other aircraft, buildings, birds, people, vehicles. Semantic segmentation models classify every pixel, creating detailed scene understanding. Real-time operation on edge GPUs enables 30+ Hz processing.

Prediction

Recurrent neural networks (RNNs) and transformers predict future trajectories of other aircraft, vehicles, and pedestrians. These predictions inform collision avoidance and path planning. Models learn from thousands of hours of real-world flight data, understanding normal vs. anomalous behavior.

Decision Making

Reinforcement learning (RL) trains policies for complex scenarios like emergency landings, abnormal weather, or system failures. Simulated training in millions of virtual scenarios creates robust decision-making unable to be programmed manually. Deep RL combines with traditional rule-based systems for safety-critical decisions.

Anomaly Detection

Unsupervised learning identifies unusual sensor readings, system behaviors, or flight dynamics indicating potential failures before they become critical. Predictive maintenance models forecast component failures based on vibration analysis, temperature trends, and usage patterns.

4.6 Safety Validation and Certification

Certifying autonomous flight systems presents unprecedented challenges. Traditional aircraft certification relies on deterministic systems with predictable behavior. Machine learning models exhibit probabilistic behavior difficult to validate exhaustively.

Simulation-Based Testing

Millions of simulated flights test autonomous systems across scenarios impossible to recreate physically—simultaneous GPS outage with sensor failure during thunderstorm, for example. Monte Carlo methods randomly vary parameters, identifying edge cases where systems perform poorly.

Wisk conducted over 1.6 million simulated autonomous flights before beginning crewed testing. Joby runs parallel real-time simulations during every test flight, comparing autonomous system decisions against multiple independent software implementations for cross-validation.

Formal Verification

Mathematical proofs demonstrate critical algorithms always behave correctly within specified bounds. Formal methods verify flight control loops maintain stability, collision avoidance always activates within required timescales, and safety-critical functions never deadlock.

Redundancy Architecture

Triple-redundant computing with diverse implementations—three independent software teams creating functionally equivalent systems on different hardware. All three systems must agree for critical decisions; disagreement triggers automatic safe mode.

4.7 Human-Machine Interface

Even fully autonomous aircraft require human oversight for high-level decisions and contingency management. Remote operations centers monitor fleet health, weather, airspace, and system status.

• Fleet Management Displays: Real-time position, status, and health of all aircraft in service
• Automated Alerting: Machine learning identifies anomalies requiring human attention
• Remote Intervention: Human operators can modify routes, delay flights, or command landings
• Passenger Communication: Video/audio links enable ground staff to assist passengers
• Emergency Override: Remote pilots can assume direct control in emergencies

4.8 Pathway to Full Autonomy

Most manufacturers pursue incremental autonomy, starting with pilot-assisted operations before progressing to full autonomy. Wisk's unique strategy targets full autonomy from day one, arguing this approach provides clearer regulatory path and better economics.

Phased Approach (Joby, Archer)

1. Phase 1 (2025-2027): Piloted operations with autopilot assistance
2. Phase 2 (2027-2030): Single pilot with autonomous backup on defined routes
3. Phase 3 (2030-2035): Full autonomy with remote pilot supervision
4. Phase 4 (2035+): Unsupervised autonomous operations

Direct Autonomy (Wisk, EHang)

Skip piloted phase entirely, pursuing autonomous certification from beginning. Requires more extensive validation but avoids incremental recertification. Chinese regulator CAAC approved EHang's autonomous operations in 2023, demonstrating feasibility.