Classical flight control theory has served aviation well for decades. PID controllers, LQR, and model predictive control (MPC) provide stable, predictable behavior when the system dynamics are well understood. But autonomous aerial vehicles increasingly operate in conditions where classical approaches alone fall short — unpredictable wind gusts, shifting payloads, degraded actuators, and environments where a mathematical model simply can’t capture every variable.
This is where machine learning enters the picture — not to replace classical control, but to augment it.
The Hybrid Approach
At Upper, our flight controller combines classical control algorithms with ML-enhanced components that handle the parts of flight control where traditional methods struggle:
Adaptive disturbance rejection. Wind is the enemy of stable flight. Classical controllers can compensate for steady winds, but turbulence and gusts introduce rapid, unpredictable forces. Our ML layer learns the vehicle’s response characteristics in real time and adjusts control outputs faster than a traditional feedback loop can react.
Payload adaptation. When a UAV picks up a payload, drops a delivery, or expends fuel, its mass distribution changes — sometimes dramatically. Our neural network continuously estimates the vehicle’s inertial properties and feeds updated parameters to the classical controller, maintaining stability without requiring manual recalibration.
Actuator degradation. Motors wear out. Propellers chip. ESCs overheat. Our system detects degraded actuators through learned performance baselines and redistributes control authority across the remaining healthy actuators — enabling continued safe operation or controlled landing.
Why Not Pure ML Control?
A fair question. End-to-end neural network controllers exist in research, but they lack the guarantees that aerospace applications demand:
- Predictability — Classical controllers have well-understood stability margins. Pure ML controllers are black boxes.
- Certification — Aviation regulators require provable safety bounds. Classical control theory provides these. ML does not, at least not yet.
- Failure modes — When a classical controller fails, it fails in understood ways. When an ML controller encounters out-of-distribution inputs, the failure mode is unpredictable.
Our hybrid architecture gives us the best of both worlds: the safety and predictability of classical control with the adaptability and learning capability of ML.
Running on Embedded Hardware
Flight control loops run at 400Hz or higher. That’s 2.5 milliseconds per control cycle — including sensor reads, state estimation, control computation, and actuator output. Our ML components are designed for this constraint:
- Lightweight neural networks with microsecond inference times
- Fixed-point arithmetic optimized for the flight controller’s onboard processor
- No cloud dependency — everything runs on the vehicle
Across Platforms
Our ML-enhanced flight controller currently supports:
- Multicopters — advanced attitude estimation and agile maneuvering in confined or GPS-denied environments
- Fixed wing — energy-optimized cruise control and autonomous waypoint navigation for long-endurance missions
- VTOL — seamless transition control between hover and forward flight with ML-managed mode switching
What’s Next
We’re expanding our ML capabilities to include predictive maintenance — using flight data patterns to forecast actuator failures before they happen, enabling proactive servicing rather than reactive repairs. We’re also investigating reinforcement learning approaches for extreme maneuvering scenarios where classical control theory has no optimal solution.
Interested in our flight controller for your aerial platform? Contact us to discuss integration and licensing options.