Computer scientists have been working on better ways to control robots using advanced algorithms. One of these techniques is called model predictive control (MPC), which involves creating a model of the robot’s movements and then using this model to plan and optimize its future actions based on a set of goals and constraints. This helps to ensure that the robot can move safely and efficiently while completing its tasks.
Computer scientists have been developing advanced algorithms to control robotic agents, such as model predictive control (MPC) techniques. MPC optimizes the agent’s behavior toward a goal while adhering to certain constraints, such as avoiding obstacles. Recently, researchers from the Technical University of Munich and University of Zurich developed Real-time Neural MPC, which integrates complex artificial neural networks (ANNs) in an MPC framework for agile robots like quadrotors.
Real-time Neural MPC builds upon the work of the University of Zurich’s Robotics and Perception Group, which developed a data-driven approach using Gaussian Processes (GPs) to boost traditional control algorithms. The researchers developed a proof-of-concept for using neural networks (Deep Learning Models) and generalized the approach to pitch the idea to the Robotics and Perception Group at the University of Zurich. From there, the two labs worked together to advance the technical work and experiments.
The Real-time Neural MPC framework combines deep learning models with online optimization of MPC. While expressive deep learning models can be computationally heavy, approximating them in real-time allows the framework to use dedicated hardware (GPUs) to process them efficiently. This allows the system to predict optimal actions for robots in real-time.
The framework allows for the combination of optimal control and deep learning while leveraging their respective optimized frameworks and computational devices. Deep learning computations can be performed on a GPU in PyTorch/Tensorflow while control optimization is performed on a CPU in compiled C code. This allows for the power of deep learning to be used in applications that were previously unachievable, such as onboard optimal control of a quadrotor.
In recent experiments, researchers from the Autonomous Aerial Systems Group at Technical University of Munich and the Robotics and Perception Group at the University of Zurich demonstrated the potential of their Real-time Neural MPC framework for controlling highly agile robots. By combining deep learning models with online optimization of model predictive control, the team achieved highly accurate tracking and reduced positional tracking errors compared to conventional MPC methods without a deep learning component.
The framework is capable of leveraging the predictive power of neural network architectures with a parametric capacity more than 4,000 times larger than previous models. By approximating these computationally heavy models in real-time, the system can predict optimal actions for robots in real-time. The researchers believe this approach could allow developers to use advanced data-driven AI techniques to better model the dynamics of robots and improve their navigation capabilities while reducing the risk of accidents.
However, the researchers also noted that one challenge of deep learning approaches is their erratic output for situations that are not part of the training data. To improve the robustness of the Real-time Neural MPC framework, the team suggested detecting these situations and providing a fallback for control to stabilize the system. As embedded systems increasingly incorporate GPU chips, the Real-time Neural MPC framework could soon play a significant role in advancing robotics research and development.
You might also be interested in reading, A robot that can autonomously explore real-world environments.
