An adaptive framework for changing-contact robot manipulation

Sidhik, Saif ORCID: 0000-0002-7031-0210 (2022). An adaptive framework for changing-contact robot manipulation. University of Birmingham. Ph.D.

Preview

Sidhik2022PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (13MB) | Preview

Abstract

Many robot manipulation tasks require the robot to make and break contact with other objects in the environment. The interaction dynamics of such tasks vary markedly before and after contact. They are also strongly influenced by the nature and physical properties of the objects involved, i.e., by factors such as type of contact, surface friction, and applied force. Many industrial assembly tasks and human manipulation tasks, e.g., peg insertion, stacking, and screwing, are instances of such `changing-contact' manipulation tasks. In such tasks, the interaction dynamics is discontinuous when the robot makes or breaks contact but smooth at other times, making it a piecewise continuous dynamical system. The discontinuities experienced by a robot during such tasks can be harmful to the robot and/or object. Designing a framework for smooth online control of changing-contact manipulation tasks is a challenging open problem.

To complete any manipulation task without data-intensive pre-training, the robot has to plan a motion trajectory, and execute this trajectory accurately and smoothly. Many methods have been developed for the former part of the problem in the form of planners that compute a suitable trajectory while considering relevant motion constraints and environmental obstacles. This thesis focuses on the relatively less-explored latter (i.e., plan execution) part of the problem in the context of changing-contact manipulation tasks. It does so by developing an adaptive, task-space, hybrid control framework that enables efficient, smooth, and accurate following of any given motion trajectory in the presence of piecewise continuous interaction dynamics. The framework makes three key contributions.

The first contribution of this thesis addresses the problem of controlling a robot performing continuous-contact tasks in the presence of smoothly-changing environment dynamics. Specifically, we provide a task-space control framework that incrementally models and predicts the end-effector wrenches, and uses the discrepancies between the predicted and measured values to revise the predictive (forward) model and to achieve smooth trajectory tracking by adapting the impedance parameters of a force-motion controller.

The second contribution of the thesis expands our framework to handle interaction dynamics that can be discontinuous due to making and breaking of contacts or due to discrete changes in the environment. We formulate the piecewise continuous interaction dynamics of the robot as a hybrid dynamical system with previously unknown discrete dynamic modes. We propose a corresponding hybrid framework that incrementally identifies new or existing modes, and adapts the parameters of the dynamics models within each such mode to provide smooth and accurate tracking of the target motion trajectory.

The third contribution of the thesis focuses on handling contact changes and reducing discontinuities in the interaction dynamics during mode transitions. Specifically, we develop a framework with a contact anticipation model that incrementally and probabilistically updates its estimates of when contact changes occur due to making or breaking contact, or changes in the properties of objects. The estimated contact positions are used to guide a transition to (and from) special `transition phase' controllers whose parameters are adapted online to minimise discontinuities (i.e., to minimise spikes in force, jerk etc) in the regions of anticipated contacts.

The stated contributions and each part of the framework are grounded and evaluated in simulation and on a physical robot performing illustrative changing-contact manipulation tasks on a tabletop. We experimentally compare our framework with some baselines to demonstrate the importance of building an incremental, adaptive framework for such tasks. In particular, we compare our controller for continuous-contact tasks with representative baselines in the adaptive control literature, and demonstrate the benefits of an incrementally-updated predictive (forward) model. We also experimentally evaluate the ability of our hybrid framework to accurately identify and model the dynamics of discrete dynamic (contact) modes, and justify the need for online updates by comparing the performance of a state of the art offline methods for hybrid dynamical systems. Finally, we evaluate the ability of our framework to accurately estimate contact positions and minimise discontinuities in the interaction dynamics in motion trajectories involving multiple contact changes.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):