Kinematic-model-free online learning for robot motion control

Abstract

Controlling robots has been a thriving field of research since their emergence. Traditionally, robots are controlled using joint space to task space transformations based on kinematic and dynamic models. However, unconventional robots like soft, morphing, malleable, transforming, and evolving ones pose challenges for accurate modeling and control. Controllers in literature typically fall into two categories: model-based and model-free. Kinematic-model-free control offers a promising solution to overcome challenges faced by conventional controllers. This method doesn’t require prior knowledge of the robot’s kinematic or dynamic models. Instead, it gathers data through exploratory actuations, building a volatile local linear model to approximate the robot’s behavior and estimate actuation signals for target movement. This thesis extends the state-of-the-art kinematic-model-free controller to tackle various problems such as tip orientation and pose control, robot configuration control, obstacle avoidance, gravity compensation, and continuous control. Initially applied to planar rigid robots, the controller was adapted for soft continuum robots, demonstrating its effectiveness through both simulation and physical experimentation. Quaternion and dual-quaternion mathematics were employed for position, orientation, and pose control. A dual-quaternionic kinematic-model-free multi-point controller was introduced to resolve redundancy in hyper-redundant manipulators, validated through simulation tests. Integration with model-predictive control enabled path planning and obstacle avoidance, validated through physical experimentation on a planar rigid robot manipulator with virtual obstacles. Modifications allowed the controller to run continuously at higher speeds, outperforming conventional methods in the face of kinematic or dynamic changes. The kinematic-model-free controller showcased versatility across different robot types, both in simulation and physical experiments. This research pushed the boundaries of the kinematicmodel- free controller and validated its effectiveness through rigorous testing.