Physics-infused Learning for Aerial Manipulator in Winds and Near-Wall Environments

This paper presents a unified control framework for aerial manipulators that integrates a physics-based blade-element model with a learning-based residual force estimator and online rotor-speed adaptation to achieve robust trajectory tracking and wall-contact operations in complex wind and near-wall environments.

The Gist

Imagine you are trying to paint a wall on the side of a skyscraper using a drone holding a paint roller. This is the dream of Aerial Manipulation: using flying robots to do physical work high up where humans can't reach.

But here's the problem: The wind is a nightmare.

When a drone flies near a tall building, the wind doesn't just blow in a straight line. It swirls, bounces off the wall, and creates chaotic pockets of air that push the drone in unpredictable ways. Traditional drones are like people trying to walk through a hurricane while holding a tray of coffee; they get knocked off course, and if they try to touch the wall, they might crash or drop their tool.

This paper presents a new "brain" for these drones that combines old-school physics with modern AI to solve this problem. Here is how it works, broken down into simple concepts:

1. The Problem: The "Guessing Game"

Most drones today use a simple rule: "If I spin my propellers faster, I go up. If I spin them slower, I go down."

• The Flaw: This works great in a calm room. But outside, near a building, the wind messes with the propellers. It's like trying to pedal a bike up a hill while someone is blowing a giant fan in your face. The simple rule fails because it doesn't understand the wind.
• The Old Solution: Scientists used to try to calculate every single air molecule (using complex math called CFD), but that takes too long for a drone to think in real-time.
• The Other Old Solution: Some tried using pure AI (Machine Learning) to guess the wind. But AI is like a student who only studied for one specific test; if the wind changes slightly, the AI gets confused and fails.

2. The Solution: The "Hybrid Chef"

The authors created a system that acts like a Master Chef who knows both the recipe (Physics) and how to improvise (AI).

Part A: The Physics Recipe (The Blade Element Model)

Think of the drone's propellers as a set of tiny airplane wings. The researchers built a "Physics Engine" that calculates exactly how the wind hits each wing.

• The Analogy: Imagine you know the exact shape of your hand and the speed of the wind. You can mathematically predict how much air will push your hand back.
• What it does: This part of the system predicts the big forces. It tells the drone, "Hey, the wind is hitting the left propeller hard; let's spin the right one a bit faster to compensate." This is the feedforward part—it acts before the drone even gets pushed off course.

Part B: The AI Improvisation (The Residual Learner)

Even the best physics math isn't perfect. There are tiny, weird swirls and vibrations the math misses.

• The Analogy: This is like a jazz musician. The sheet music (Physics) tells you the main notes, but the musician (AI) listens to the room and adds the little "flourishes" to make it sound right.
• What it does: The AI looks at what the Physics model missed. It learns the "leftover" errors. If the Physics model says "push left," but the drone still drifts right, the AI learns, "Oh, there's a hidden gust I didn't see. I'll add a little extra push to the right."

Part C: The Adaptive Coach (The Online Observer)

Wind changes constantly. A gust might hit one second and vanish the next.

• The Analogy: Imagine a coach standing next to the drone. If the drone starts to wobble, the coach instantly shouts, "Adjust your balance!"
• What it does: This system constantly updates itself in real-time. If the wind gets stronger than expected, the "Coach" instantly recalibrates the AI and the Physics model to handle the new intensity.

3. The "Rotor Speed" Trick

Usually, when a drone needs to move, it just tells the motors: "Spin at 50% power."

• The Innovation: This paper changes the instruction to: "Spin at the exact speed needed to fight this specific wind."
• The Analogy: Instead of just telling a car "Go faster," the system calculates exactly how much gas to press based on the slope of the hill and the wind resistance. It allocates power to each propeller individually to cancel out the wind's push before it even happens.

4. The Results: Flying in the Storm

The researchers tested this in a simulation where a drone had to fly in a figure-eight pattern near a wall and even touch the wall to "paint" it, all while strong winds blew.

• The "No-Brain" Drone: Got blown away immediately when the wind got strong.
• The "Pure AI" Drone: Did okay in mild wind but got confused when the wind changed unexpectedly.
• The "Hybrid" Drone (This Paper):
  • It stayed on track even in extreme winds (12 m/s, which is a strong gale).
  • It could touch the wall gently without bouncing off, even when the wind was trying to push it away.
  • It was the only one that didn't crash when the wind conditions were totally new and unseen.

The Big Picture

This paper is about giving drones common sense.
Instead of just following a rigid rulebook or blindly guessing with AI, the drone now understands the physics of the air (the recipe) and has the experience to adapt to the weird stuff (the improvisation).

This means that in the future, we might see drones that can safely clean skyscrapers, fix power lines, or plant seeds in forests, even when the weather is terrible, because they finally learned how to "dance" with the wind instead of fighting it.

Technical Summary

Here is a detailed technical summary of the paper "Physics-infused Learning for Aerial Manipulator in Winds and Near-Wall Environments."

1. Problem Statement

Aerial Manipulation (AM) extends UAV capabilities from passive observation to physical interaction (e.g., bridge painting, power line de-icing). However, current AM systems are primarily validated in controlled indoor environments and struggle in real-world scenarios due to two critical aerodynamic challenges:

• Nonlinear Wind Disturbances: Strong ambient winds create uncertain forces that simplified thrust models (e.g., thrust $\propto \omega^2$) fail to capture accurately.
• Proximity-Induced Flow Effects: Operating near infrastructure (walls, buildings) creates complex, non-uniform flow fields (wall effects, ground effects) that distort local airflow. These effects cause nonlinear interactions that are difficult to model analytically but are computationally too expensive for real-time control if using high-fidelity Computational Fluid Dynamics (CFD).
• Generalization Gap: Purely learning-based models (data-driven) often fail to generalize to unseen wind conditions or coupled dynamics during physical contact with surfaces.

The core problem is developing a control framework that is computationally efficient for real-time use, robust to strong and varying winds, and capable of maintaining precise contact with vertical structures in aerodynamically complex environments.

2. Methodology

The authors propose a hybrid physics-infused learning framework that integrates a physics-based model with a data-driven residual estimator and an online adaptive observer. The approach consists of four main components:

A. Physics-Based Blade Element Model (BEMT)

Instead of using simplified quadratic thrust models, the authors employ Blade Element Momentum Theory (BEMT) to model rotor aerodynamics.

• Mechanism: BEMT calculates aerodynamic forces and moments by integrating lift and drag along the propeller blade span, accounting for local flow conditions (freestream wind, induced velocity, and vehicle motion).
• Parameter Learning: The aerodynamic coefficients of the BEMT model (lift curve slope, drag scale, stall angle) are learned from free-flight data to create a "nominal" physics-based model. This provides a structured, interpretable feedforward estimate of aerodynamic forces.
• Rotor Speed Allocation: A novel allocation strategy solves for the specific rotor speeds required to generate desired thrust and moments based on the learned BEMT model, rather than a static inverse map. This accounts for wind-induced variations in rotor efficiency.

B. Learning-Based Residual Estimator

To capture aerodynamic effects not covered by the BEMT model (e.g., body drag, sensor biases, complex flow interactions), a neural network is trained to predict the "residual" force.

• Architecture: A multi-head residual network with a shared backbone.
• Inputs: Vehicle twist (linear/angular velocity), pose (quaternion), rotor speeds, and local freestream wind velocity.
• Output: The residual aerodynamic force vector ($\mathbf{F}_{res}$) that the BEMT model failed to predict.
• Training: The network minimizes the difference between the BEMT-predicted force and the force reconstructed from sensor measurements (IMU + motion capture).

C. Online Adaptive Disturbance Observer

To handle time-varying discrepancies and unmodeled contact forces during wall interaction, an adaptive observer updates a correction term ($\hat{\boldsymbol{\epsilon}}$) in real-time.

• This component uses a covariance update law (similar to a Kalman filter structure) to estimate and compensate for the remaining mismatch between the predicted and actual forces, ensuring robustness against drift and sudden environmental changes.

D. Unified Control Framework

The total disturbance estimate ($\hat{\mathbf{F}}_d$) used in the controller is the sum of three terms:
$$\hat{\mathbf{F}}_d = \mathbf{F}_{phys} + \mathbf{F}_{res} + \hat{\boldsymbol{\epsilon}}$$
Where $\mathbf{F}_{phys}$ is the lateral component of the BEMT prediction, $\mathbf{F}_{res}$ is the neural network output, and $\hat{\boldsymbol{\epsilon}}$ is the adaptive correction.

3. Key Contributions

1. Hybrid Modeling Framework: Integration of a physics-based BEMT model with a learning-based residual estimator. This leverages the interpretability and generalization of physics while using data to correct unmodeled complexities.
2. Wind-Aware Rotor Allocation: A control allocation strategy that uses the learned BEMT model to dynamically adjust rotor speeds based on wind conditions, reducing thrust mismatch and unintended lateral moments.
3. Near-Wall Simulation Environment: Construction of a high-fidelity simulation environment featuring spatially varying wind fields near a vertical wall and a deformable end-effector for contact tasks.
4. Robustness to Unseen Conditions: Demonstration that the hybrid approach generalizes better to extreme wind conditions (outside the training distribution) compared to purely data-driven methods.

4. Results

The framework was evaluated in a simulated environment with a quadrotor performing two tasks: a figure-eight trajectory (free flight) and a circular trajectory with wall contact. Tests were conducted under wind speeds of -4, -8, and -12 m/s (applied in X and Z directions).

• Disturbance Estimation: The BEMT-residual model achieved significantly lower prediction loss compared to a zero-wind baseline, proving that the physics prior captures the majority of aerodynamic effects, allowing the neural network to focus on fine-grained errors.
• Tracking Performance (Figure-Eight):
  • At -4 m/s and -8 m/s, the proposed method performed comparably to a state-of-the-art data-driven baseline (DAIML).
  • At -12 m/s (unseen extreme condition), the uncompensated controller failed completely. The DAIML baseline showed significant tracking errors due to poor generalization. The proposed method maintained the lowest RMS tracking error in both X and Z directions.
• Wall-Contact Performance:
  • In the circular wall-contact task, the proposed method maintained stable contact force and trajectory tracking even at -12 m/s.
  • The DAIML baseline and uncompensated controllers exhibited significant trajectory deformation and unstable contact forces at high wind speeds.
  • The proposed method successfully mitigated the combined effects of strong wind and contact-induced disturbances, keeping the vehicle on the reference path with minimal lateral drift.

5. Significance

This paper addresses a critical bottleneck in deploying aerial manipulators for real-world infrastructure maintenance. By moving beyond simplified thrust models and purely data-driven black boxes, the proposed physics-infused learning approach offers a balanced solution:

• Reliability: It ensures robustness in extreme, unseen wind conditions where pure learning methods fail.
• Efficiency: It avoids the computational cost of real-time CFD by using a pre-learned physics model.
• Precision: It enables stable physical contact with structures in aerodynamically complex zones (near walls), which is essential for tasks like painting, inspection, and de-icing.

The results suggest that combining structured physical knowledge with flexible data-driven correction is the most promising path forward for autonomous aerial systems operating in unstructured, dynamic environments.