Deep Learning a Quadrotor Dynamic Model for Multi-Step Prediction

Mohajerin, Nima; Mozifian, Melissa; Waslander, Steven L.

doi:10.1109/icra.2018.8460840

Cited by 35 publications

(25 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Due to their ability to identify patterns in large amounts of data, deep neural networks represent a promising approach to model aerodynamic effects precisely and computationally efficient. A recent line of works employs deep neural networks to learn quadrotor dynamics model purely from data, for both continuous time formulations [8,9] as well as discretetime formulations [15,16,38,39]. While approaches relying entirely on learning-based methods have high representative power and the potential to also learn complex interaction effects, they require large amounts of data to train and careful regularization to avoid overfitting.…”

Section: Related Workmentioning

confidence: 99%

“…Thus, learning a high-order dynamics model requires a method for regression of a nonlinear function in a high-dimensional input space. Deep neural networks have shown to excel at such highdimensional regression tasks and have already been applied to dynamic system modeling [8,9,[15][16][17]. Despite showing promising performance, such purely-learned models require large amounts of data and require careful regularization to avoid overfitting.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

NeuroBEM: Hybrid Aerodynamic Quadrotor Model

Bauersfeld

Kaufmann

Foehn

et al. 2021

Robotics: Science and Systems XVII

View full text Add to dashboard Cite

Fig. 1: Long-exposure images depicting quadrotor trajectory tracking at speeds up to 65 km/h in a large-scale motion-capture system. The captured data is used to fit a hybrid quadrotor model combining blade-element-momentum (BEM) theory with a neural network compensating residual dynamics. This hybrid model reproduces the flown trajectories in simulation with a positional RMSE error reduction of over 50% compared to state-of-the-art.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

NeuroBEM: Hybrid Aerodynamic Quadrotor Model

Bauersfeld

Kaufmann

Foehn

et al. 2021

Robotics: Science and Systems XVII

View full text Add to dashboard Cite

show abstract

“…That is, specific relations within a model shall be learned on the basis of reference data while keeping physically established relations unchanged. This is different from pure black box modeling (Mohajerin et al, 2018). The idea of localized adaption of single equations is especially applicable to object-oriented modeling (like in Modelica) and aims at keeping learning results understandable to the user by separation from existing "white box" relations.…”

Section: Artificial Neural Network For Model Calibration and Augmentmentioning

confidence: 99%

The OpenModelica Integrated Environment for Modeling, Simulation, and Model-Based Development

Fritzson¹,

Pop²,

Abdelhak³

et al. 2020

MIC

106

View full text Add to dashboard Cite

OpenModelica is a unique large-scale integrated open-source Modelica-and FMI-based modeling, simulation, optimization, model-based analysis and development environment. Moreover, the OpenModelica environment provides a number of facilities such as debugging; optimization; visualization and 3D animation; web-based model editing and simulation; scripting from Modelica, Python, Julia, and Matlab; efficient simulation and co-simulation of FMI-based models; compilation for embedded systems; Modelica-UML integration; requirement verification; and generation of parallel code for multi-core architectures. The environment is based on the equation-based object-oriented Modelica language and currently uses the MetaModelica extended version of Modelica for its model compiler implementation. This overview paper gives an up-to-date description of the capabilities of the system, short overviews of used open source symbolic and numeric algorithms with pointers to published literature, tool integration aspects, some lessons learned, and the main vision behind its development.

show abstract

“…With the limitations of the traditional models, dataintensive methods like deep learning adopt self-supervision in data collection, thus allows creation of large dataset and achieves better outcomes [14], and recent works utilizes deep feed-forward neural network or deep Long-Short-Term-Memory (LSTM) neural network for several control problems like quadrotor [15], robot-assisted dressing [16] and under-actuated legged millirobots [17]. These works achieves better outcomes than the traditional models of model-based RL.…”

Section: Related Work a Model-based Rlmentioning

confidence: 99%

Learning to Guide: Guidance Law Based on Deep Meta-Learning and Model Predictive Path Integral Control

et al. 2019

View full text Add to dashboard Cite

In this paper, we present a novel guidance scheme based on model-based deep reinforcement learning (RL) technique. With model-based deep RL method, a deep neural network is trained as a predictive model of guidance dynamics which is incorporated into a model predictive path integral (MPPI) control framework. However, the traditional MPPI framework assumes the actual environment similar to the training dataset for the deep neural network which is impractical in practice with different maneuvering of the target, other perturbations, and actuator failures. To address this problem, our method utilizes meta-learning technique to make the deep neural dynamics model adapt to such changes online. With this approach, we can alleviate the performance deterioration of standard MPPI control caused by the difference between the actual environment and training data. Then, a novel guidance law for a varying velocity interceptor intercepting maneuvering target with desired terminal impact angle under actuator failure is constructed based on the aforementioned techniques. The simulation and experiment results under different cases show the effectiveness and robustness of the proposed guidance law in achieving successful interceptions of maneuvering target. INDEX TERMS Missile guidance, model predictive control, meta-learning, deep reinforcement learning, impact angle constraint.

show abstract

Deep Learning a Quadrotor Dynamic Model for Multi-Step Prediction

Cited by 35 publications

References 12 publications

NeuroBEM: Hybrid Aerodynamic Quadrotor Model

NeuroBEM: Hybrid Aerodynamic Quadrotor Model

The OpenModelica Integrated Environment for Modeling, Simulation, and Model-Based Development

Learning to Guide: Guidance Law Based on Deep Meta-Learning and Model Predictive Path Integral Control

Contact Info

Product

Resources

About