Contact Transition Control: An Experimental Study

Successful control of contact transitions is an important capability of dextrous robotic manipulators. In this paper we examine several methods for controlling the transition from free motion to constrained motion, with an emphasis on minimizing fingertip load oscillations during the transition. A new approach, based on input command shaping, is discussed and compared with several methods developed in prior research. The various techniques were evaluated on a one-axis impact testbed, and we present results from those experiments. The input shaping method was found to be comparable, and in some cases superior, to existing techniques of contact transition control. Introduction Contact transitions can be notoriously difficult to control. These events involve all of the well known problems of force control, with the added difficulties of non-zero approach velocities and discontinuous dynamic characteristics. If a robotic manipulator is to interact effectively with its environment, however, it must frequently make and break contact with foreign objects. An event-driven, reflexive capability to manage grasp state transitions is crucial for successful, practical dextrous manipulation. Compliant fingertips have been used to eliminate some contact transition problems, but many difficult tasks remain. We find that even with soft fingertips, contact transitions tend to excite instability and/or result in undesirably high impact forces. These phenomena can be particularly bothersome when manipulating with sensorladen fingertips, devices that are sensitive to "glitches" in force control. Our particular goal in this work is to achieve smooth, stable transitions between motion and force control. We want to avoid instability and large force spikes during the controller transition, while increasing the grasp force from zero to the desired level as rapidly as possible. Contact problems have not been neglected in prior research. Mills and Lokhorst [5] implemented a discontinuous controller that sought to reduce the problems of contact instability and also managed events of contact loss and trajectory tracking. Hogan [2] used impedance control in experiments involving contact transitions. His implementation achieved stability against a stiff environment and avoided controller transitions and inverse kinematic computations. Vossoughi and Donath [9] also employed impedance methods in tests with varying environment stiffnesses. Actively compliant fingertips containing electrorheological fluids were explored by Akella, Siegwart, and Cutkosky [1], who managed to control the damping characteristics of a fluid-filled fingertip to minimize force oscillations following the onset of contact. Youcef-Toumi and Gutz [9] developed a dimensionless representation of impact behaviour and used integral force compensation with velocity feedback to improve impact response. Khatib and Burdick [4] proposed a method for dissipating impact oscillations that involved increasing the velocity gains of a proportional-derivative force controller for a limited time following impact. By disabling the high velocity feedback gain after the impact oscillations decayed, response times to subsequent force commands were decreased. More recently, Qian and De Schutter [6] presented an active nonlinear damping approach. This method examined the force signal derivative, and effectively added a coulomb friction term to the output force command when this derivative exceeded a threshold. We will examine some of these methods, and compare them to a relatively new technique: input command preshaping. This method has enjoyed success in many applications of position control, as shown in Singer [7] and Hyde [3], but to the authors' knowledge, preshaping has never been applied to force control problems. Input command preshaping is essentially a feedforward technique that employs linear system theory and a plant's inherent dynamic characteristics to suppress vibration. Shaping an input is straightforward. First, the system frequency targeted for suppression must be identified to within about 10% (similar to the identification of a mode targeted for control by a dominant 2nd order controller). This mode is input into closed-form equations used to calculate a train of impulses, a process which can occur off-line if necessary. The impulses are convolved with the input command in real time, imposing only a small computational burden and producing a shaped input which is then fed to the plant. At runtime, the requirements are that the servo rate and actuator authority be sufficient to accurately produce the desired commanded force function -in practice, servo rates and actuator authorities found adequate for stable force control will be more than sufficient. When applying the shaped input, the system output will be largely devoid of vibrations at the targeted frequency. As an example, a standard single mode shaper has three impulses. When a step input is convolved with this impulse train, the modified input resembles a three-step staircase, as shown in Figure 1. Any commanded trajectory (step, ramp, or sinusoid; force, position, etc.) can be "shaped" with this method. Preshaping produces only small system response delays, and is relatively insensitive to variations in plant parameters. Since the method operates outside of a standard control loop, preshaping can be used in conjunction with any of the other feedback methods mentioned above. Shapers can also be used to eliminate multiple-mode vibration. For full details on input command preshaping, see Singer [7] and Hyde [3]. -0.2 0 0.2 0.4 0.6 0.8 1 1.2