The 1.55-Kilogram Rocket That Could Democratize Aerospace Control

Imagine a rocket standing on a launchpad, its engines roaring, when a sudden gust of wind slams into the vehicle. The guidance computer has to make split-second decisions—tilt the thrust, fight the disturbance, keep the payload pointed skyward. It's one of the hardest control problems in aerospace engineering, and testing new algorithms to handle it means risking millions of dollars of hardware. Now, researchers have built something unusual to solve this problem: a flying robot that looks like a bottle rocket mounted on top of a drone. It weighs just 1.55 kilograms, costs a few thousand dollars, and can crash a hundred times without consequence—yet captures the same essential physics that a real launch vehicle faces.

The platform, called QuadRocket, is the subject of a new paper published in IEEE Transactions on Aerospace and Electronic Systems by Pedro Santos, Joel Reis, Paulo Oliveira, and Carlos Silvestre (Santos et al., 2026). What makes it noteworthy isn't just what it can do, but how it does it: by treating the quadrotor as a thrust-vector actuator, the researchers achieved almost global trajectory tracking—the kind of robust, rock-solid stability that real rockets need but that most laboratory testbeds struggle to deliver. It's a bridge between the clean mathematics of control theory and the messy reality of aerospace engineering.

The Science

Launch vehicles are notoriously difficult to control. The thrust from the engines doesn't point straight down through the center of mass; instead, it's deliberately tilted to steer the rocket, creating a coupling between where the force is applied and how the vehicle rotates. This is called thrust-vector control, and it's what allows a rocket to pitch, yaw, and roll on command. But the same physics that makes this steering possible also makes the system unstable: tilt the thrust wrong, and the rocket will flip end over end. Testing new control strategies—ones that might be more efficient, more robust, or better at handling disturbances—has traditionally required either expensive simulations or real hardware that costs more to replace than most universities can afford.

The researchers' starting point was simple: if a quadrotor can provide force and torque in arbitrary directions, could it serve as a surrogate engine for a rocket-like vehicle? The answer, it turns out, is yes—but with a twist. Most prior approaches used the quadrotor as the main body with a small pendulum attached. Here, the researchers flipped the arrangement. Their cylindrical carbon-fiber body, weighing 0.64 kilograms, is actually heavier than the quadrotor itself at 0.45 kilograms. This matters enormously: in real rockets, the payload and fuel tank dominate the mass, not the engine. By making the attached structure the largest contributor to the system, the QuadRocket better mimics the dynamics of an actual launch vehicle.

The system works like this: a carbon-fiber cylinder sits atop the quadrotor through an aluminum universal joint, which allows the body to tilt up to 40 degrees in any direction while keeping the quadrotor's rotors roughly level. The battery that powers the quadrotor is mounted near the top of the cylinder, shifting the center of mass upward and increasing the lever arm for control. This creates what's known as a flying inverted pendulum—a configuration where a heavy mass on top of a mobile base tries constantly to tip over, and the control system must continuously counteract this instability. The researchers describe it as "physically analogous to a rocket controlled via thrust-vectoring," because the net thrust force is applied along a direction not fixed to the main body axis, generating the same coupled translational and rotational dynamics.

For control design, the team made two key modeling choices. First, they abstracted away the joint constraints and modeled the system as a single axisymmetric rigid body—essentially pretending the quadrotor is just providing a vectored force along the rocket's longitudinal axis. Second, they exploited the vehicle's symmetry by using a reduced-attitude representation on the two-sphere. This mathematical trick separates the pitching and rolling of the rocket's nose from its rotation around its own axis (yaw), which is exactly what you want when your control objective is to point the vehicle somewhere, not to rotate it to a particular heading.

The control architecture breaks into two loops working in concert. The outer loop computes the desired thrust-vector direction to track a given trajectory, using an adaptive backstepping controller that can handle unknown constant disturbances—think of these as wind gusts or modeling errors that the system hasn't been told about. The inner loop commands the quadrotor's angular velocity so that it actually achieves that thrust direction, treating the rotors as the actuator that generates the force. A control-point transformation moves the reference point from the center of mass to the center of oscillation, canceling out a phenomenon called non-minimum-phase behavior that would otherwise make the system sluggish or unstable.

The experiments took place in an indoor motion-capture arena, where infrared cameras track reflective markers with millimeter precision. This allowed the researchers to measure the vehicle's position and orientation in real time and feed that data back to their controllers. The quadrotor's flight computer ran BetaFlight software in manual mode, receiving thrust and angular velocity commands directly from the control algorithm via radio link.

What They Found

The paper reports three main categories of results: simulation validation, indoor flight testing, and disturbance compensation. In simulation, the researchers tested trajectory tracking performance using a figure-eight reference path. The control system drove all tracking errors to zero with asymptotic convergence—the theoretical gold standard meaning the vehicle actually reaches and maintains the desired trajectory rather than just getting close to it. The error metrics they define cover position, velocity, thrust-vector direction, and angular velocity, and all of them settle to near-zero values after an initial transient phase.

For the experimental validation, the motion-capture system provided ground-truth measurements at 200 Hz. The results show position tracking errors well below 10 centimeters in steady state, with the vehicle following commanded trajectories accurately even as the pendulum body swings and tilts. The key insight is that despite the complexity of the coupled dynamics—the quadrotor and rocket body influence each other continuously—the unified control design keeps everything stable.

The mass distribution of the system is central to understanding its behavior. The carbon-fiber structure contributes 41.3 percent of the total mass, the quadrotor adds 29 percent, the battery and power cables account for 22.6 percent, and the universal joint and shaft make up the remaining 7.1 percent. At a total mass of 1.55 kilograms, this puts the center of mass well above the quadrotor, creating the large moment of inertia that makes the system behave like a real rocket rather than a toy.

QuadRocket Mass Distribution by Component

Label	Value
Carbon Fiber Structure	0.64 kg
Quadrotor	0.45 kg
Battery & Cables	0.35 kg
Universal Joint & Shaft	0.11 kg

The carbon fiber body dominates the mass budget at 41.3%, making the vehicle dynamically similar to real rockets where payload mass exceeds engine mass.

The adaptive element of the controller proved its worth when the researchers introduced unknown disturbances. By estimating and compensating for constant exogenous forces—simulating effects like wind or unmodeled friction—the system maintained tracking performance where a non-adaptive controller would have degraded. This capability is crucial for real-world deployment, where no model is perfect and conditions are never exactly as predicted.

The backstepping design, which constructs the controller step by step by treating outputs from one stage as virtual control inputs for the next, yielded a single Lyapunov function that encapsulates all tracking errors. This mathematical artifact serves as a certificate of stability: if the function always decreases, the system cannot become unstable. The researchers show that their particular construction achieves almost-global convergence, meaning the controller works for every possible initial condition except a set of measure zero—essentially everywhere in practice.

Mass Share of Each Component

Label	Value
Carbon Fiber Structure	41.3
Quadrotor	29
Battery & Cables	22.6
Universal Joint & Shaft	7.1

Unlike typical quadrotor testbeds where the drone dominates, here the payload-like body represents the largest mass fraction—a key feature for rocket-like dynamics.

The dynamic-surface-based inner-loop controller handles the fact that the quadrotor has its own actuation dynamics—its rotors can't change speed instantly—and avoids the mathematical麻烦 of differentiating noisy signals. This is a practical concern: real sensors have noise, and taking derivatives amplifies that noise. By using a filter to estimate derivatives instead, the controller remains robust to measurement error.

Why This Changes Things

The significance of this work isn't primarily about building another drone. It's about democratizing access to a class of control problems that have traditionally required either massive budgets or enormous risk tolerance. Thrust-vector control is how rockets steer; it's how spacecraft land on Mars; it's how missiles hit moving targets. But the people developing new algorithms to do this more efficiently or more robustly have had limited options for rapid, iterative testing.

Existing testbeds, like those developed at Swiss research institutions, use gimbaled propellers mounted on drones—a mechanical linkage that physically tilts the thrust direction. The QuadRocket approach is different: instead of tilting the propeller, you tilt the entire vehicle. This avoids mechanical constraints like servo rate limits and backlash, and it leverages the mature, well-understood control infrastructure that already exists for quadrotors. If you can write an attitude controller for a quadrotor, you already have half the solution.

The researchers also see applications beyond rocket testing. The cylindrical body could carry cargo, making the platform useful for autonomous delivery in confined spaces. It could inspect ceilings and other civil infrastructure, where its ability to position a sensor package at various angles while maintaining stability is valuable. These secondary uses suggest that the underlying technology—accurate thrust-vector control for asymmetric payloads—has broad applicability in robotics.

From a control-theory perspective, the paper contributes a rare example of almost-global stability for a thrust-vector-controlled system with unknown disturbances. The rocket guidance literature often relies on separate position and attitude loops with linear approximations, which are easier to analyze but less robust to real-world conditions. By merging these loops and using nonlinear techniques, the researchers demonstrate that global guarantees are achievable even when the system has underactuation and external disturbances.

The fact that the heaviest part of the vehicle is not the propulsion system is worth emphasizing. In most quadrotor-based testbeds, the quadrotor dominates the dynamics; the attached payload is an afterthought. Here, the rocket body is the primary inertial element, which means the control strategies developed have to contend with significant inertia and coupling effects that wouldn't exist in a traditional configuration. This makes the testbed more realistic and the results more transferable to actual aerospace applications.

What's Next

The paper identifies several open questions and avenues for future work. The current platform operates indoors under motion-capture supervision, which provides precise state estimation but limits where experiments can run. Outdoor testing with GPS or visual-inertial localization would expand the operational envelope, though wind and other disturbances would add complexity. The researchers also note that their current disturbance model assumes constant or slowly varying forces; more sophisticated models for gusty wind or flexible-body dynamics could increase robustness.

On the control side, the almost-global stability result comes with a caveat: there exist initial conditions where the system may fail to converge. In practice, these conditions are measure-zero sets that the vehicle is unlikely to encounter by chance, but for safety-critical applications, a globally stable controller—guaranteed to work from any starting point—might be preferable even if it sacrifices some performance. The trade-off between almost-global and global stability is a recurring theme in geometric control, and the QuadRocket provides a testbed to explore it.

The researchers also envision extending the platform to simulate additional launch-vehicle phenomena. Previous work by some of the authors has used quadrotors with attached inverted pendulums to represent aerodynamic instability and fuel sloshing. Adding these effects to the QuadRocket could make it an even more comprehensive testbed for guidance and control algorithms, capturing not just the rigid-body dynamics but also the vibrational modes that plague real rockets.

There are practical considerations too. The current platform uses off-the-shelf components and 3D-printed parts, making it relatively inexpensive and easy to reproduce. But it requires an expensive motion-capture system for state estimation. Developing visual-inertial estimation algorithms that work on a single onboard computer would reduce cost and complexity, bringing the platform within reach of more laboratories and potentially even commercial applications.

The broader question is what happens when control algorithms proven on platforms like the QuadRocket make their way into actual aerospace systems. The path from laboratory demonstration to flight certification is long and full of regulatory hurdles. But testbeds like this lower the barrier to developing and debugging new algorithms, which could eventually accelerate the pace of innovation in launch-vehicle guidance. Whether that's through cheaper access to space, more reliable rockets, or entirely new mission profiles, the implications extend well beyond the 1.55 kilograms of carbon fiber and aluminum hovering in a University of Macau laboratory.