⚡ electrical engineering

A Mathematical Simulation Framework for Comparative Evaluation of Rocket Control Strategies: From Classical PID to Hybrid Adaptive Architectures

This paper introduces an open-source web-based mathematical simulation framework that systematically benchmarks nineteen rocket control strategies against complex physical models, revealing that while PID-based architectures generally outperform others, specific hybrid variants like Impulse control can significantly improve fuel efficiency during landing.

Original authors: Lucas Lima Freitag

Published 2026-07-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lucas Lima Freitag

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a very tall, very wobbly rocket ship. Your job is to blast off into the sky and then, the tricky part, come back down and land gently on a tiny target without tipping over or running out of fuel. It's like trying to balance a broomstick on your hand while someone is blowing wind at you and shaking the floor.

To figure out the best way to steer this rocket, a researcher named Lucas built a giant, super-detailed video game in a web browser. But instead of just playing, this game is a mathematical simulation. It's a digital playground where you can test 19 different "autopilot" brains to see which one is the best pilot.

The Digital Test Track

The simulation is incredibly realistic. It doesn't just move a dot on a screen; it calculates how the rocket's weight changes as it burns fuel, how the air gets thinner as it goes up, and even how the rocket shakes like a guitar string when the engine fires. It adds fake wind, turbulence, and even "noise" to the sensors (like static on a radio) to make sure the autopilots aren't just lucky.

The researcher ran two main scenarios:

The Launch: Starting from the ground and shooting up to 800 meters.
The Landing: Dropping from 500 meters down to a soft stop at 0 meters.

The 19 Contenders

The paper tested a huge lineup of control strategies. Some were the old-school classics, like PID (a very common type of controller that acts like a thermostat, constantly checking the temperature and adjusting the heat). Others were fancy new ideas, like Fuzzy Logic (which makes decisions based on "maybe" and "sort of" rather than strict math), Neural Networks (trying to mimic a brain), and Model Predictive Control (which tries to calculate the future steps ahead of time).

There were also some clever new twists the researcher invented, like Impulse Control (which treats the engine like a firework, firing in short, powerful bursts and then coasting) and Smart-PID (which changes its rules depending on how close it is to the target).

The Big Results: Who Won?

After running thousands of simulations, here is what the data showed:

1. The Old Guard Still Rules (Mostly)
The PID controllers were the consistent champions. In the launch scenario, they scored between 81 and 82 out of 100. In the landing scenario, they scored 74 to 75 out of 100.
Why did they win? The paper suggests it's because they use a "cascade" structure. Think of it like a team: one part of the brain decides where you want to go, a second part decides how fast to get there, and a third part actually pushes the throttle. This separation of duties makes them very stable and good at ignoring the wind.

2. The Fuzzy Logic Surprise
In the launch phase, a Fuzzy Logic controller actually edged out the PID, scoring 82 compared to PID's 81. It was slightly better at ignoring the wind disturbances, but only by a tiny margin.

3. The Fuel Saver
Here is the coolest finding for the landing phase. The Impulse Control strategy (the "firework" approach) managed to land the rocket using 12.7% less fuel than the standard PID. It landed in about 44.8 seconds with a fuel burn of 123.8 kg, compared to the PID's 141.8 kg. This proves that sometimes, firing the engine in short, smart bursts is better than a steady burn.

4. The "Too Fancy" Failures
The paper explicitly rules out some high-tech ideas for this specific job in this simulation.

Model Predictive Control (MPC): This was the "super-optimizer" that tried to calculate the future. While it was fast at the start, it failed catastrophically during the landing. It scored only 41 out of 100 and caused the rocket to tip over (pitch-over) because it got confused by the sensor noise.
Neural Networks: The "brain" controller, which used random weights without being trained, failed to provide reliable control.
Adaptive SMC: A controller that tried to change its own rules in real-time failed miserably in the noisy environment, dropping its noise rejection score to 0%.

What This Means

The paper doesn't claim to have solved rocket science forever. Instead, it shows that in a controlled, simulated environment, simple, well-tuned math often beats complex, futuristic math. The "smart" algorithms that look great on paper struggled when the simulation added real-world messiness like wind and sensor static.

However, the study did prove that you can get better results by tweaking the simple rules. The Impulse controller showed that you can save a significant amount of fuel (over 12%) just by changing how you fire the engine, not by inventing a new type of brain.

The researcher made all the code, the math models, and the results available for anyone to check. This means the "game" is open for anyone to try their own ideas, ensuring that future rocket pilots can be tested rigorously before they ever leave the ground.

Technical Summary: A Mathematical Simulation Framework for Comparative Evaluation of Rocket Control Strategies

Problem Statement
The reusability of launch vehicles, exemplified by Vertical Take-Off and Vertical Landing (VTVL) systems like the SpaceX Falcon 9, relies heavily on solving complex control problems during the powered descent phase. This phase requires simultaneous regulation of altitude, velocity, and attitude under rapidly changing mass properties, nonlinear aerodynamics, atmospheric disturbances, and sensor noise, all within strict fuel constraints. While numerous control strategies exist in literature—ranging from classical PID and LQR to Sliding Mode Control (SMC), Model Predictive Control (MPC), and reinforcement learning—comparative studies are often limited. Most existing research evaluates only two or three methods within a single framework, making it difficult to understand relative strengths and weaknesses under identical physical conditions. Furthermore, hardware-in-the-loop testing is often impractical for systematic, large-scale comparisons.

Methodology
The authors developed a mathematical simulation framework implemented as an open-source web platform using JavaScript. The core of the framework is a high-fidelity physics engine that models six-degree-of-freedom (6-DOF) rigid-body dynamics (simplified here to planar 3-DOF translational and 1-DOF rotational dynamics) using a fourth-order Runge-Kutta (RK4) integrator with a fixed time step of 0.005 s.

The simulation environment incorporates realistic physical models, including:

Atmosphere: International Standard Atmosphere (ISA) up to 11 km.
Disturbances: Wind modeled via power-law shear and stochastic gusts, turbulence generated via a von Kármán spectral model approximated by a Dryden filter, and structural vibrations (first three bending modes).
Sensor Noise: Zero-mean Gaussian perturbations applied to position, velocity, attitude, and rate sensors.

The framework utilizes a modular architecture where control models are injected via a standardized functional interface $u(t) = f(\tilde{x}(t), r(t), \Delta t)$ , allowing any mathematical control law to be tested without modifying the physics engine. Two mission scenarios were simulated: a launch to 800 m and a landing from 500 m, both subject to identical environmental conditions (6 m/s wind base, turbulence intensity 1.0).

Key Contributions
The paper makes four primary contributions:

Open-Source Framework: A browser-executable simulation platform that models 6-DOF dynamics, realistic atmospheric disturbances, and sensor noise, enabling reproducible benchmarking without external dependencies.
Novel Controller Variants: Implementation of ten novel controller variants beyond standard implementations, including Adapt-SMC (adaptive gain sliding mode), Auto-Iter (phase-based iterative learning), Fuel-Saver (four-zone coasting PID), Hybrid-IL (noise prediction with iterative correction), Impulse (bang-bang with coast phases), Neural (random-weight feedforward), Noise-Pred (AR(4) cascade), RL-HC (reinforcement heuristic), Short-Act (quantized deadzone), and Smart-PID (zone-dependent integral management).
Systematic Benchmarking: A comparative evaluation of 19 distinct control strategies across launch and landing phases using seven quantitative metrics: settling time, RMS tracking error, mean absolute attitude, fuel consumption, noise rejection, disturbance rejection, and a composite overall score.
Data Transparency: Provision of complete source code, all control model implementations, and raw trajectory data as supplementary material.

Results
The evaluation yielded the following key findings:

PID Dominance: PID-based architectures consistently achieved the highest composite scores in both scenarios (81–82/100 for launch, 74–75/100 for landing). The authors attribute this to integral action eliminating steady-state error and cascade structures providing natural saturation handling.
Fuzzy Logic Performance: In the launch scenario, the Fuzzy Logic controller achieved the highest overall score (82/100), marginally outperforming PID (81/100) due to superior disturbance rejection (48.8% vs. 44.5%).
Fuel Efficiency: The "Impulse" controller (a bang-bang strategy with adaptive pulse timing) achieved the best fuel economy in the landing scenario, consuming 123.8 kg compared to 141.8 kg for standard PID—a 12.7% reduction—while maintaining comparable settling times and RMS errors.
Limitations of Advanced Methods:
- Adapt-SMC: Showed poor performance in noisy environments, with noise rejection dropping to 0% because the adaptation law amplified sensor noise.
- MPC: Achieved the fastest initial ascent but suffered catastrophic pitch-over during landing due to a lack of attitude regularization in the fixed-horizon optimization.
- Neural Networks: A randomly initialized feedforward network provided basic control but did not outperform classical methods.

Significance
The paper claims that a well-structured mathematical simulation framework enables systematic comparisons that are impractical with physical testing alone. By decoupling the physics engine from the control interface, the framework facilitates rapid prototyping and allows researchers to evaluate controllers spanning seven orders of magnitude in computational complexity under identical conditions. The study highlights that while advanced theoretical methods (like MPC and neural networks) offer mathematical guarantees, they may lack the practical robustness of simpler, well-tuned PID variants in the presence of sensor noise and turbulence. The authors posit that the framework serves as a reproducible benchmark for VTVL control algorithm development, bridging the gap between theoretical advances and practical application through transparent scientific validation.

The Digital Test Track

The 19 Contenders

The Big Results: Who Won?

What This Means

More like this