Examenspresentation av Lasse Kötz och Jonathan Almgren

Examiner: Jonas Sjöberg

Abstract:
State estimation is an important aspect in a large number of robotic applications. With recent advancements within the field of Machine Learning (ML), it has become increasingly interesting to study how Neural Networks (NNs) can be applied to overcome this problem.

This master thesis consists of implementing, training and comparing three different NN architectures that, given a video stream as input, estimate the current leaning angle of a Wheeled Inverted Pendulum (WIP) system. This is done to such precision and inference speed that it can be used to control an unstable real time system. To deploy and demonstrate the real time performance of these ML-models, a self-balancing robot was constructed. The demonstration robot consists of a custom designed platform with 3D-printed components mounted together by threaded rods and actuated by two brushed DC-motor equipped with encoders. The core processing is performed on a Raspberry Pi 3B boosted by a Tensor Processing Unit (TPU) that helps with processing the incoming camera data through the NNs.
Control of the system is performed using two cascade Proportional–Integral–Derivative controllers (PID), where one outer loop controls the horizontal speed of the robot while the inner loop controls the leaning angle of the robot. Design of the control system is performed using classic control methods and its’ biggest challenge is handling the slower sampling rate of camera data compared to the alternative solution of using an IMU for angle estimation.

The models showed results with Mean Absolute Errors (MAE) reaching as far down
as 0.8255° and a standard deviation of 0.4072 in ideal cases. Through signal processing, this could be reduced further under certain conditions. When running on the Raspberry Pi 3B hardware, the deployed NN reached a sampling rate of 60 Hz, which was too slow to get accurate performance in controlling the system. Simplified test runs were conducted on upgraded hardware which allowed it to reach adequate sampling rates for stability but could not be deployed on the physical robot due to project limitations.

Analysis of test run data shows that ML-models have a tendency to predict conser-
vatively for leaning angles of higher magnitude. Through signal processing methods the prediction error can be reduced slightly for certain cases at the cost of reduced performance in other cases. Due to the nature of the demonstration platform, which shouldbalance around low leaning angles, the processing is optimised around these cases.

Welcome!

Lasse, Jonathan and Jonas