Control of Multiple Robotic Manipulators Background J.T. Wen Rensselaer Polytechnic Institute Ken Kreutz-Delgado University of California, San Diego
Automatica, Volume 28, Issue 4, July 1992
The purpose of the paper is to present a unified
perspective on the motion and force control problem for multiple-arm systems and to propose a framework for stability analysis of those systems. Design Philosophy Assume the arms, the payload, and the grasps are all rigid. This assumption holds for most robotic systems. This implies that the composite contact force from all the arms can be decomposed into two orthogonal components, only one of which effects the motion of the object. The other force is internal to the held object: the squeeze force Therefore, as they are orthogonal, force control does not affect object motion, although object motion does affect the internal force The authors then present the design philosophy for these systems: design a stable motion control law without consideration of force control. Then design a stable force control law by treating the inertial force as a perturbation (which is independent of the force control). Control Architectures (Motion Control) The choice of the variable considered in the control input results in different control architectures. 1. The acceleration of a generalized coordinate. The feedback-linearization approach 2. The composite tip force of all of the arms. The arms-as-actuators approach 3. The joint torques of all of the arms. The full dynamics approach Model for Multiple-Arm Systems Assume arms and held object are rigid and the object is also rigid No relative motion at the contact A squeeze force can be applied without any motion
Equations of motion are:
Model for Multiple-Arm Systems
These equations can be combined to solve for the contact force f as:
f is uniquely solvable if and only if [ J A] has full
rank. Under some general conditions, [ J A] having full rank at every kinematically feasible configuration is a general property, so we assume f is uniquely solvable. Model for Multiple-Arm Systems For a given composite tip force f, there exists a unique orthogonal decomposition for fc , the force on the object:
This is called the move/squeeze decomposition
As a result, any term in of the form , with the last component in the squeeze subspace does not affect motion of the system i.e., if the force only squeezes the object, it does not effect the motion of the object. The control torque can then be decomposed into
where the compensation is model-based.
General Procedure 1. Design a stable motion controller for , ignoring the squeeze force. 2. Design the squeeze force controller by regarding the motion-induced component of the squeeze force as a disturbance that is unaffected by the squeeze force or the squeeze control. Motion Control Feedback linearization and arms-as-actuators both require full dynamical model information of the manipulator. For various reasons, the authors focus on the full dynamics approach For single arm systems, much work had been done at the time of writing The authors extend this to multi-arm systems in this paper Full Dynamics Approach The Lyapunov function candidate is chosen to be the total kinetic energy (for the arms and payload) and an artificial potential energy which has its global minimum at the set point. There are three types of variables that can be used in the potential energy: 1. Joint position (of all arms) 2. Tip position (of all arms) 3. Generalized coordinates The first two over-specify the configuration of the system and are not generalized coordinates. For generalized coordinates, there are many possible choices. Set Point Control Problem Consider the following Lyapunov function candidate
where U* is any of the following:
Set Point Control Problem The derivative of V along the solution is therefore:
respectively for the different values of U*.
By the invariance principle, this implies the asymptotic convergence of to zero, i.e., the arms always reach a steady state configuration. This holds true even for infeasible joint and tip set points in the joint and tip level feedback cases. Set Point Control Problem For the joint and tip level feedback cases, the solution manifolds may contain a configuration where the motion control component of each arm works against each other and does not produce any motion. These are known as the joint and tip Jam Manifolds, respectively. While the fixed set point control laws are a useful exercise, especially in how it points out issues unique to multiple-arm control (such as the Jam Manifold), the fixed set point control paradigm is fundamentally flawed because the closed loop trajectory transient is not controlled. For initial condition far away from the desired set point, the transient is typically so wild that these control laws are virtually unusable. Moving Set Points Consider desired trajectories that belong to
These are mild restrictions in point-to-point
operations---the desired trajectories are those that reach a steady state. We call S the class of moving set points. Moving Set Points Result 1: If the desired trajectory of the multiple- arm system belongs to S, then the system with control laws in (13) and PD feedback as in Table 3 has a known stability property (Table 4). If the initial tracking error is 0, the maximum trajectory tracking error is inversely proportional to the size of the PD gains.
For the generalized coordinate feedback case, the
Jam Manifold is guaranteed to be just the zero configuration. For the other two cases, it may be that the nonzero error jammed configuration exists (not discussed in this paper). Moving Set Points If full model information is assumed, precise force control at every moment in time is possible: while the model-independent control law reduces finite force error only with the use of high gains. Squeeze Force Control As to not disturb the motion control, the squeeze torque is chosen to be of the form
where is a full rank matrix whose columns
span the null space of The control variable for force control is now , and the squeeze projection of the contact force can be written as:
where is the motion-induced squeeze force
This tends to 0 if an asymptotically stable motion control law is used. Squeeze Force Control can be chosen in one of two ways: 1. An optimization problem 2. A set point problem The authors mostly ignore the optimization problem, referring to previous literature, and turn their attention to the set point problem. If it is desired to have tight control of the squeeze force, such as in the manipulation of delicate objects, it may be necessary to choose from a set point squeeze control problem. However, without using full model information, the squeeze force can only be controlled asymptotically. Squeeze Force Control An important property of the motion-induced squeeze force is that it is not effected by the squeeze force, and can therefore be treated as an external disturbance. Let fcs,des be the desired value of the squeeze force Assuming a stable motion control strategy, we are interested in: Stability of fcs Transient performance (maximum force error) Convergence rate Noise reduction Robustness with respect to time delay in force measurements Squeeze Force Control For asymptotic stability, a feedforward control will suffice, but provides no noise reduction. If arm tip forces are measured, then a feedback strategy is preferable, due to the hope for added noise reduction and improved transient performance. The rigidity assumptions made necessitate extra care in the control design. Recognizing that the problem is caused by the algebraic loop due to the proportional force feedback, the authors suggest pre-processing the measured force by a strictly causal filter before feedback, making the control law: Squeeze Force Control This feedback law clearly follows the stability property. For transient behavior, the filter must be analyzed in the Laplace domain. Essentially, all control objectives are satisfied with this law, provided that it is a stable filter with a pole at the origin. For example, an integrator has the desired properties. Result 2. For the multiple-arm control system under consideration, if the arm configuration converges to a steady state (i.e. velocity converges to zero) and the gravity load is fully compensated, then either the feedforward controller or the feedback controller with C a Jamming Under non-generalized-coordinate feedback (such as faint level and tip level control laws), the arms converge to the Jam Manifold. The members in the Jam Manifold other than the desired set point are called the jammed configurations, since they represent balance of all the internal forces due to the applied torque and gravity. Essentially, if a quantity related to the kinematics of the robot is sufficiently small, the map describing the motion of the robot becomes a contraction with the only fixed point at the desired state. Simulation Results Performance comparison between fixed and moving set point control laws. Integral feedback squeeze control vs feedforward squeeze control. Simulation Results (1) Simulation Results (2)
In general, feedback force control yields smaller
overshoot, much better transient response, and no steady state error (due to the integral action). Conclusions The rigidity assumption on the held object allows a useful decomposition of the contact spatial force into a component moving the held object and another component contributing only to the internal forces. The motion and force control problems can be independently analyzed. In the motion control, depending on the choice of variables used in the control design, one obtains three different control structures: the full dynamics, arms-as-actuator and feedback linearization control structures. Conclusions The actuator redundancy in a multiple-arm control system can be resolved either through an optimization (load balancing) or a squeeze force set point control problem. Integral force feedback control can achieve good performance and possesses good robustness properties. Future research topics include a flexible held object and/or a flexible environment (in the contact case) and robustness analysis with respect to the grasp map. Questions?