Physics-Informed Neural Network for Multirotor Slung Load Systems Modeling

Gil Serrano,Marcelo Jacinto,Jose Ribeiro-Gomes,Joao Pinto,Bruno J. Guerreiro,Alexandre Bernardino,Rita Cunha

2024-05-15

Abstract:Recent advances in aerial robotics have enabled the use of multirotor vehicles for autonomous payload transportation. Resorting only to classical methods to reliably model a quadrotor carrying a cable-slung load poses significant challenges. On the other hand, purely data-driven learning methods do not comply by design with the problem's physical constraints, especially in states that are not densely represented in training data. In this work, we explore the use of physics informed neural networks to learn an end-to-end model of the multirotor-slung-load system and, at a given time, estimate a sequence of the future system states. An LSTM encoder decoder with an attention mechanism is used to capture the dynamics of the system. To guarantee the cohesiveness between the multiple predicted states of the system, we propose the use of a physics-based term in the loss function, which includes a discretized physical model derived from first principles together with slack variables that allow for a small mismatch between expected and predicted values. To train the model, a dataset using a real-world quadrotor carrying a slung load was curated and is made available. Prediction results are presented and corroborate the feasibility of the approach. The proposed method outperforms both the first principles physical model and a comparable neural network model trained without the physics regularization proposed.

Robotics

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to establish a dynamic model capable of accurately predicting the future state of the system in a multi - rotor UAV with a suspended - load system. Traditional modeling techniques based on classical methods face significant challenges when dealing with such complex systems, while pure data - driven methods can capture the dynamic characteristics of the system but are difficult to meet physical constraints, especially in states where the training data is not fully covered. Therefore, the paper proposes a method combined with Physics - Informed Neural Network (PINN), aiming to improve the prediction accuracy and generalization ability of the model by fusing prior physical knowledge and data - driven learning methods. Specifically, the main contributions of the paper are as follows: 1. **Propose an encoder - decoder architecture**: This architecture utilizes LSTM (Long - Short - Term Memory network) and the attention mechanism to predict the future state sequence of the multi - rotor UAV with a suspended - load system. At the same time, relaxation variables and physics - based loss terms are introduced as regularization terms to limit the search space and ensure that the prediction results conform to physical laws. 2. **Construct a new real - world dataset**: This dataset is obtained through 48 independent flight operations in the Motion Capture (MOCAP) site and is used to train the proposed model. Through this method, the paper not only improves the accuracy of the model in short - term prediction but also shows the good performance of the model beyond the prediction time range considered during training, verifying the effectiveness and practicality of the proposed method.

Physics-Informed Neural Network for Multirotor Slung Load Systems Modeling

Neural Predictor for Flight Control with Payload

Adaptive Multi-Surface Sliding Mode Control with Radial Basis Function Neural Networks and Reinforcement Learning for Multirotor Slung Load Systems

Physics-Inspired Temporal Learning of Quadrotor Dynamics for Accurate Model Predictive Trajectory Tracking

Physics-Infused Machine Learning Based Prediction of VTOL Aerodynamics with Sparse Datasets

HPA-MPC: Hybrid Perception-Aware Nonlinear Model Predictive Control for Quadrotors with Suspended Loads

Physics-informed multi-LSTM networks for metamodeling of nonlinear structures

NeuroBEM: Hybrid Aerodynamic Quadrotor Model

A Physics-Informed Neural Network Approach to Augmented Dynamics Visual Servoing of Multirotors

Physics-informed Neural Networks to Model and Control Robots: a Theoretical and Experimental Investigation

Learning Long-Horizon Predictions for Quadrotor Dynamics

Neural Network Based Model Predictive Control for a Quadrotor UAV

Prediction of the Flight Dynamics of Maneuvering Multirotor Aircraft

A novel residual hybrid dynamic model for unmanned helicopters: combining both physical and deep learning models

A Computationally Efficient Learning-Based Model Predictive Control for Multirotors under Aerodynamic Disturbances

Physics Encoded Blocks in Residual Neural Network Architectures for Digital Twin Models

Real-time Neural-MPC: Deep Learning Model Predictive Control for Quadrotors and Agile Robotic Platforms

Intelligent Position Controller for Unmanned Aerial Vehicles (UAV) Based on Supervised Deep Learning

An Improved Method for Swing State Estimation in Multirotor Slung Load Applications

Back to Newton's Laws: Learning Vision-based Agile Flight via Differentiable Physics

Flying Quadrotors in Tight Formations using Learning-based Model Predictive Control