Integration of planning and execution in force controlled compliant motion

Abstract

This paper presents the Compliant Task Generator, a new approach for the automatic conversion of a geometrical contact path into a force-based task specification. A contact path planner generates a sequence of six-dimensional poses and corresponding contact formations, while a hybrid robot controller expects a desired wrench, twist and the local wrench and twist subspaces. Our approach automatically converts a geometrical path description into a force based tasks specification for the hybrid controller, based on a user specified input of the magnitudes and the norms of the desired contact force and execution speed. The approach applies to all contact motions between known polyhedral objects, and is verified in real world experiments.

Introduction

Compliant motion tasks are manipulation tasks that involve contacts between the object manipulated by a robot and the environment in which it operates. To cope with uncertainties these operations must be carried out using passive or active force control. Indeed small errors in the object models can give rise to high interaction forces. Whereas passive force control relies on compliance elements placed at the wrist of the robot, in active force control the robot controller modifies the trajectory depending on the forces arising during the interaction [13].

To specify force controlled robot tasks M.T. Mason introduced the Task Frame Formalism (TFF) [29] an intuitive and manipulator independent formalism for the specification of force controlled robot tasks in the Hybrid Control Paradigm (HCP) [34]. Bruyninckx and De Schutter made an extensive catalogue of TFF models and specifications [3]. While the TFF is useful to specify many elementary contact tasks [20], it cannot cope with more complex operations involving multiple simultaneous contacts, even in the case of polyhedral objects; the TFF is limited to translational and rotational components along the same axes, orthogonal to each other. Recently, De Schutter presented a constraint-based task specification framework that overcomes the limitations of the TFF, and provides a powerful interface to specify complex compliant motion tasks involving multiple (contact) constraints [12]. However, these approaches themselves have no “intelligence” such as planning or advanced sensor processing capabilities, but fully rely on human intervention and intuition to specify a task compatible with the framework. Indeed, the programmer not only has to specify the task but also has to foresee the input sensor signal to control it. For more complex tasks, involving multiple contacts and changes in contact formations, this can prove to be extremely difficult.

Different approaches exist to automatically generate the desired sequence of contact formations in such a case. Programming by Human Demonstration [6], [16], [31], [35] gathers wrench, pose and contact data about a task, while a human demonstrates the task, in a virtual or real environment using for instance a haptic device or a demonstration tool. A different approach, as used in this paper, involves a geometrical planner that automatically generates a path in the contact space of the manipulated object and its environment based on their geometrical models. Xiao and Ji developed such a compliant planner [18]. Due to the complexity of the problem this planner finds such a path in two stages. In the first one the planner automatically generates the contact state space between two arbitrary polyhedral objects in terms of a contact state graph [38]. Even between two simple polyhedral objects, hundreds of contact formations are possible. In this graph the possible contact states and their adjacency relations are represented. In the second stage the planner generates a compliant path by compliant interpolation between neighbouring contact states. However, this path only contains geometrical and topological information. For the complete specification of the compliant task the forces to be applied to the environment along the path have to be described.

In this paper we present an approach to automatically generate a task specification for a hybrid controller, based on the output of the geometrical planner, called the Compliant Task Generator, in which the user specifies the desired magnitudes for the contact wrench and the manipulator twist. This allows a complex task plan to be generated based on the known geometrical models of the objects that can be generated on a real robot manipulator under active force control without expert user intervention. The method has been implemented and real world experiments have been carried out to validate it. This work is complementary to–and can be integrated with–previous work of our research group, which focused on the identification of contact states and the estimation of geometrical parameters using iterative stochastic estimation tools such as Kalman filters or particle filters [16], [23], [31]. To improve the identification and estimation, in [25] the active sensing problem is formulated and decoupling it into smaller optimization problems.

The remainder of the paper is organized as follows: Section 2 briefly reviews the concept of contact formations and the contact state graph. Section 3 discusses the output of the compliant planner and the input to the hybrid controller, in other words the planner and controller primitives. Section 4 describes the automatic generation of the controller primitives, using the planner primitives and information about the desired contact force level and execution speed. Section 5 explains the internals of the hybrid controller. Section 6 discusses the invariance and the robustness of the method. The experimental setup and the obtained results are discussed in Section 7. Finally, Section 8 contains the conclusions together with future extensions and improvements of the method.