Henrik Hose: Safe Neural Network Controller for Agile Robots
Fast feedback responses, stability, and constraint satisfaction are critical requirements for control in robotics to ensure safety.
Model predictive control (MPC) achieves stability and constraint satisfaction, but is notoriously slow to evaluate.
Approximation of such MPC controllers via (deep) neural networks (NNs) allows for fast online evaluation.
However, the approximation introduces inaccuracies that can cause instabilities or constraint violations.
In this project, novel methods for offline validation and safe online evaluation of approximations of MPC type controllers are developed.
This work builds upon existing results in statistical offline validation, online safety certification in control, and explores the use of formal verification methods.
Novel approximate MPC schemes with offline validation and safe online evaluation methods are evaluated in real-world problems from the robotics domain, such as the Wheelbot.
The Wheelbot, a small reaction wheel balancing robot, was originally developed at the DSME and MPI Stuttgart under the supervision of Prof. S. Trimpe.
A video of the Wheelbot is available here.
The Wheelbot is a challenging robotics test bed for non-linear control when balancing, and even hybrid-systems with contact switches for stand-up maneuver.
The next generation --- the Mini Wheelbot --- is engineered for production in small fleet quantities to serve as a hardware test bed at DSME.