In recent years, computer scientists have created increasingly sophisticated algorithms to direct the movements of robotic agents. These include model predictive control (MPC) strategies that use a model of the agent's dynamics to optimize its approaching behavior towards a given target while satisfying a set of constraints (for example, not hitting obstacles).
Model Predictive Control Strategies and Artificial Neural Network Technologies
Real-Time Neural MPC is a framework that integrates complex model architectures based on artificial neural networks (ANNs) into an MPC framework for mobile robots (i.e. quadrotors-drones). It was recently developed by researchers at the Technical University of Munich and the University of Zurich. This idea, reported in IEEE Robotics and Automation Letters, extends an idea previously created by the University of Zurich Robotics and Perception Group.
Tim Salzmann and Markus Ryll Tech, researchers at the Autonomous Air Systems Group at the Technical University of Munich, said: “We came across the excellent work of the Robotics and Sensing Group led by Davide Scaramuzza to develop their core idea of having data-driven (learned) components that power 'traditional' control algorithms. We were immediately captivated.
“After developing a proof-of-concept to extend their approach using Gaussian Processes (GPs) to general Neural Networks (Deep Learning Models), we presented our idea to the Robotics and Perception Group at the University of Zurich. From this point on, the technical work and testing of the two laboratories progressed jointly and ignited a new partnership.”
Deep learning models and online MPC optimization are combined in the new framework proposed by Salzmann, Ryll et al. Deep learning expression models require a lot of computation. Still, the framework can use specialized hardware (GPUs) to effectively render these models online in real-time. This allows their systems to predict in real time the best course of action for the robots.
Salzmann and Ryll said: “The Real-Time Neural MPC framework enables the two domains to combine optimal control and deep learning, while allowing both parts to take advantage of their own highly optimized frameworks and computational devices. “So, control optimization can be performed in C code compiled on the CPU, while deep learning computations can be performed on a GPU in PyTorch/Tensorflow. This enables deep learning to be used in applications that were hitherto impractical, such as on-board quadrotor optimal control.”
Researchers evaluate their framework through a series of simulations and field-based tests. In these studies, it is especially used to control the movements of a highly mobile quadrotor in real time.
The ability to use neural network topologies with parametric capacity 4.000 times higher than those previously used to regulate the motions of mobile robots in real time allowed them to achieve very promising results. They also discovered that the framework they developed can reduce spatial tracking errors by up to 82% compared to traditional MPC approaches without a deep learning component.
According to Salzmann and Ryll, “in robotics, we look for meaningful patterns of the dynamics of controlled systems and their interaction with the environment (eg aerodynamic effects, tire friction, etc.)”. “Although these are often difficult to analyze, learning-based methods, especially those using neural networks, can capture dynamics and interaction effects. However, the accuracy of the model increases with the size of the neural network. When deep learning models are used in real-time neural MPC, the model is much more powerful and efficient in predictive control than was previously possible.”
GPU chips are slowly making their way into embedded systems, as demonstrated by the recently introduced Nvidia Jetson platform. This team of researchers has soon developed a framework that will allow designers to leverage the high predictive power of sophisticated data-driven AI techniques to better model the dynamics and interactions of robots with the environment, reducing the risk of accidents and improving navigation capabilities, integrating GPU chips.
Salzmann and Ryll noted that there are many unexplored possibilities for further study. “The output of deep learning methods can be unpredictable for situations not included in the training data (Non-Distributed OOD). Robustness in OOD conditions will come from detecting those conditions and providing a fallback for control to stabilize the system.”
Günceleme: 13/03/2023 14:09