Design, Model Identification and Control of a Low-Cost Wheeled Mobile Robot Using Differentiable Physics
Wednesday, May 27, 2020, 03:00pm - 04:30pm
Speaker: Yanshi Luo
Location : Remote via Webex
Prof. Mridul Aanjaneya, Prof. Abdeslam Boularias, Prof. He Zhu and Prof. Santosh Nagarakatte
Event Type: Qualifying Exam
Abstract: With the availability of affordable 3D printers and single-board microcontrollers such as Arduino and Raspberry Pi, there is growing interest in building affordable robots for various tasks. We designed a low-cost wheeled mobile robot and proposed an analytical model for predicting its motion that is derived from first principles. Further, we showed how the robot motion can be accurately predicted under the influence of motor torques and friction forces by learning unknown physical parameters using gradient descent. By combining our analytical model with a neural network, our robot is able to automatically predict the control signal sequences for driving itself along predefined curves. Compared to existing black-box system identification methods and other data-driven techniques, our model is more explanatory and requires much less data. Through experimental evaluation, the robot has great alignment with the predicted motion with a limited amount of data.