Learning and Control

1. Learning-Based Control for Underactuated Balance Robots

We present a Gaussian process (GP)-based tracking control of underactuated balance robots in which an actuated subsystem is required to follow a desired trajectory, while an unactuated, unstable subsystem needs to be kept balanced. The GP models are used to capture the coupling effects between the actuated/unactuated subsystems through a constructed balance equilibrium manifold (BEM). Optimization-based algorithm is used to obtain the BEM estimation. The control design takes advantage of the structural property of the robot dynamics and is built on the GP models with a data selection algorithm. Stability analysis is given to guarantee the tracking control performance. The control design and comparison with other controllers are demonstrated through experiments on a rotary pendulum.

Quanser Rotary Inverted Pendulum

Step flow for GP-based control

Nominal model selection
Data collection with real robot
Offline GP model training and validation
Controller design
Implementation

Data Collection for GP-based Modeling

Control with Learned GP Models

Robust Performance with GP-based Control

Related Work

F. Han and J. Yi, “Stable Learning-Based Tracking Control of Underactuated Balance Robots,” in IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 1543-1550, April 2021. Concurrent Submission to 2021 ICRA Conference. PDF, Video1, Video2, Video3
F. Han and J. Yi, “On the Learned Balance Manifold of Underactuated Balance Robots,” 2023 IEEE International Conference on Robotics and Automation (ICRA), London, United Kingdom, 2023, pp. 12254-12260. PDF
F. Han and J. Yi, “Learning-based Safe Motion Control of Vehicle Ski-Stunt Maneuvers,” 2022 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), Sapporo, Japan, 2022, pp. 724-729. PDF
C. Deng, Y. Gong, F. Han, S. Liao, J. Yi and B. Yuan, “VLSI Hardware Architecture for Gaussian Process,” 2020 54th Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, USA, 2020, pp. 121-124. PDF

Page updated

Google Sites

Report abuse