An autonomous robot bicycle for active safety tests

Licentiatavhandling, 2024

With the rapid emergence of autonomous driving technology, safety tests have become crucial for validating the functionality of these systems. High-fidelity scenarios of the tests typically involve various types of road users. To ensure safety during tests, human drivers are excluded, and instead, dummies and robots are used to simulate real road users. The test objects should follow pre-defined trajectories timely, and interact with each other. Then the onboard ADAS systems' reaction may be tested in a controllable and repeatable manner.
In this thesis, we developed an autonomous robot bicycle to serve in the aforementioned safety tests. To build the robot, an electric bicycle has been modified with sensors and actuators, and the target is to follow a specified trajectory with a dummy cyclist mounted on the saddle.

Several modules are essential for this purpose. Firstly, a state estimator is designed to compute the bicycle's states using sensor fusion techniques. Secondly, a speed-dependent controller is developed through system identification technique, which estimates the dynamics of the unstable bicycle based on experimental data. Finally, an Iterative Learning Controller (ILC) is designed to exploit the repetitive nature of safety tests, enhancing trajectory tracking performance through iteration. These modules are detailed in the three papers included in this thesis.

In the first paper on state estimation, we developed a Kalman-based stationary filter to estimation the bike position, speed, heading, roll and steering angles, which are crucial for balancing and trajectory tracking. The estimator integrates data from 4 sensors: an IMU, a steering encoder, a longitudinal speed sensor and GPS, along with predictions results from physical models. To evaluate this estimator, we integrated it into the control loop, and the experimental results were consistent with our observation.

The second paper addresses the design of a speed-dependent controller through gray-box identification. To balance the bicycle in a large speed range, a speed-dependent gain-scheduling controller is developed based on an identified bicycle model. We proposed a method to identify such a semi-linear model, which is nonlinear only with respect to longitudinal speeds. This method leverages knowledge of the bicycle's physical model structure and collects data in a closed-loop. Experimental results demonstrate that the controller can extrapolate the dynamics beyond known speed configurations.

The third paper discusses the iterative learning control technique for trajectory tracking. The ILC evaluates longitudinal and lateral errors during trajectory tracking experiments and computes time and position-dependent feedforward signals to compensate for errors and unmodeled dynamics. Simulation results show that errors are reduced in the both dimensions, improving tracking performance through repetition.