⚡ electrical engineering

Flatness-Preserving Residual Learning for Real-Time Tight Quadrotor Formation Flight

This paper proposes a physics-informed residual learning framework that preserves differential flatness to enable computationally efficient, real-time tight quadrotor formation flight, achieving significantly reduced tracking errors and matching state-of-the-art nonlinear model predictive control performance with an order of magnitude less computation using minimal training data.

Original authors: Pei-An Hsieh, Fengjun Yang, Nikolai Matni, M. Ani Hsieh

Published 2026-07-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pei-An Hsieh, Fengjun Yang, Nikolai Matni, M. Ani Hsieh

Original paper licensed under CC BY 4.0 (http://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 a flock of tiny, buzzing drones trying to fly in a super-tight huddle, like a swarm of bees trying to squeeze through a keyhole. The problem? When they get that close, the air gets messy. The drone on top pushes air down (like a leaf blower), which hits the drone below and throws it off course. If the drones don't know this is happening, they might crash into each other or spin out of control.

For a long time, the smartest way to handle this was to use a super-brainy computer program called "Nonlinear Model Predictive Control" (NMPC). Think of NMPC as a genius chess player who calculates every single possible future move for the next hour before making a move. It works great, but it's so heavy and slow that it needs a massive computer to run. Most small drones don't have computers that big.

The Big Idea: The "Flat" Shortcut
The researchers in this paper came up with a clever trick. They realized that if they taught the drone's brain a specific type of lesson—one that strictly follows the rules of physics regarding position and speed—they could keep the math "flat."

In the world of drone physics, "flat" doesn't mean lying on the ground; it means the math stays simple enough that you don't need a supercomputer to solve it. It's like realizing that instead of calculating the entire weather system to know if you need an umbrella, you can just look at the sky right above your head. By forcing their learning model to only look at where the drones are and how fast they are moving (and nothing else), they kept the math simple.

The Hybrid Brain: Physics + Learning
To make this work, they gave the drones a "hybrid brain."

The Physics Part: They started with a basic, hand-written rule about how air flows from one drone to another (like knowing that hot air rises). This is the "physics prior."
The Learning Part: Then, they let a neural network learn the messy, weird details that the basic rule missed.

Because they built the learning part on top of the physics rules, the drone could predict exactly how the wind from its neighbor would push it. This allowed them to use a much simpler controller (like a standard cruise control) that could instantly cancel out the wind push before it even happened.

The Results: Fast, Light, and Accurate
The team tested this on real hardware—actual tiny drones flying through a small window. Here is what they found:

Less Mistakes: Their method reduced the average tracking errors by 31% compared to the standard way of flying without this extra help.
Speed: It matched the performance of the heavy, slow NMPC "genius" controller but required 7 times less computing power.
Training: They only needed 30 seconds of flight data to teach the system how to handle the wind.
Loop Rate: The system could make decisions and adjust the motors 5 milliseconds at a time (which is incredibly fast).

What They Didn't Do (and Why It Matters)
The paper is very careful to say what this is not. They explicitly ruled out using the "genius" NMPC approach for these small drones because it's too heavy. They also didn't just guess that this would work; they proved it with real hardware experiments, not just computer simulations.

They also didn't try to teach the drone to learn everything about the air. If they had let the drone learn every possible variable, the math would have broken, and the "flatness" would have vanished, making the simple controller useless. By sticking to the specific rule that the wind depends only on position and speed, they kept the system stable and fast.

The Bottom Line
This isn't a magic wand that solves every flying problem forever, but it is a major step forward. It shows that you can get high-fidelity, crash-free flying for tight formations without needing a supercomputer on board. It's like teaching a race car driver to anticipate a turn by feeling the road, rather than having a satellite calculate every inch of the track for them. The result? A drone swarm that can fly tight, fast, and safely, even with a tiny brain.

Technical Summary: Flatness-Preserving Residual Learning for Real-Time Tight Quadrotor Formation Flight

Problem Statement

Flying multiple quadrotors in tight proximity introduces severe challenges due to turbulent aerodynamic interactions, such as downwash. If these unmodeled disturbances are ignored, they can degrade tracking performance or lead to catastrophic collisions. While learning residual dynamics from flight data offers a solution, integrating general, unstructured residual models into control pipelines typically requires Nonlinear Model Predictive Control (NMPC). However, NMPC is computationally expensive and scales poorly with model complexity, making it unsuitable for resource-constrained flight stacks requiring real-time performance (e.g., 5ms loop rates). Existing methods that impose structural constraints (e.g., smoothness) often lack the expressivity needed for complex multi-agent interactions or fail to provide anticipatory feedforward compensation.

Methodology

The authors propose a physics-informed residual dynamics learning framework that balances model fidelity with control efficiency by ensuring the joint multi-quadrotor system remains differentially flat.

1. Flatness-Preserving Parameterization

The core theoretical contribution is restricting the learned residual dynamics to depend strictly on the joint positions and velocities of the quadrotor team ( $\xi = [p_1, v_1, \dots, p_N, v_N]^\top$ ).

Differential Flatness: By limiting the residual dependency to kinematic states, the augmented dynamics preserve the differential flatness property of the nominal rigid-body system. This allows the system states and control inputs to be reconstructed algebraically from flat outputs (positions and yaw) and their derivatives.
Controller Design: Leveraging this preserved flatness, the authors design a computationally efficient feedback linearization controller.
- Feedforward Compensation: The controller uses the learned residual to cancel aerodynamic disturbances via feedforward terms in both thrust and torque calculations.
- Linear Tuning: Because the system is flattened, the feedback gains can be tuned using standard linear control techniques (e.g., PID), avoiding the need for complex nonlinear optimization.
- Robustness: To avoid noise amplification from high-order derivatives (specifically the second derivative of the residual), the torque input utilizes a high-gain proportional controller on angular velocity error rather than exact dynamic extension.

2. Physics-Informed Residual Learning

To mitigate potential generalization loss caused by the restricted expressivity of the flatness-preserving parameterization, the residual model is structured as a sum of a physics-based Reduced-Order Model (ROM) and a learned neural network term:
$\hat{f}(\xi) = \hat{f}_d(h_0) + \hat{f}_\Theta(h_0)$

Physics Prior ( $\hat{f}_d$ ): A pairwise downwash model adapted from prior work, which estimates induced airflow velocity based on relative vertical ( $\Delta z$ ) and radial ( $\Delta r$ ) separation. This provides a physically grounded foundation.
Learned Term ( $\hat{f}_\Theta$ ): A neural network that learns the residual errors not captured by the ROM, operating on rotationally equivariant features derived from the relative kinematics.

Key Contributions

The paper outlines four primary contributions:

Theoretical Proof of Joint Flatness: The authors demonstrate that augmenting nominal rigid-body dynamics with a structured residual parameterization (dependent only on joint positions and velocities) preserves the joint differential flatness of a multi-quadrotor team. They explicitly derive the mappings to reconstruct states and inputs from flat outputs.
Sample-Efficient Learning Framework: A residual learning framework is proposed that utilizes known physical structures (symmetries and ROMs) to enable learning with limited data.
Efficient Controller Design: A controller is designed that leverages the learned residuals for predictive feedforward compensation while maintaining the computational benefits of flatness-based control.
Experimental Validation: The method is evaluated in both simulation and hardware experiments, demonstrating performance comparable to state-of-the-art NMPC baselines with significantly reduced computational requirements.

Results

Tracking Performance: Hardware experiments show that the proposed framework reduces average tracking errors by 31% compared to nominal baselines that ignore aerodynamic effects.
Comparison to NMPC: The proposed controller achieves tracking performance on par with state-of-the-art NMPC baselines (specifically the KNODE-MPC framework) but requires an order of magnitude less computation (specifically 7 $\times$ less compute in the identical hardware setup).
Data Efficiency: The system achieves stable, tight formation flight (merging from 1.1m to 0.1m separation) using under 30 seconds of training data.
Real-Time Capability: The approach operates successfully at a 5ms loop rate, unlocking high-fidelity aerodynamic compensation for compute-constrained flight stacks.

Significance and Claims

The paper claims to be the first to successfully apply feedback linearization controllers with learned residuals to tight quadrotor formation flight.

The significance of this work lies in its ability to bridge the gap between high-fidelity aerodynamic modeling and real-time computational constraints. By preserving differential flatness, the authors enable the use of simple, linear control tuning techniques for complex, learned multi-agent dynamics. This approach eliminates the computational bottleneck of NMPC while maintaining its performance, thereby enabling stable, high-speed formation flight in environments where aerodynamic interactions are critical. The authors position this as a solution for "compute-constrained flight stacks," demonstrating that stable tight formation is achievable with minimal training data and low-latency execution.