Torque Distribution and Slip Minimization in an Omnidirectional Mobile Base Yuan Ping Li, Denny Oetomo, Marcelo H. Ang Jr.* National University of Singapore 1 ent Ridge Crescent, Singapore 1196 *mpeangh@nus.edu.sg Chee Wang Lim Singapore Institute of Manufacturing Technology 71 Nanyang Drv, Singapore 6875 Abstract Two forward kinematic models which are used in the control of an omnidirectional mobile base are evaluated. These two models result in different sensitivities to joint position error. The analysis and experimental results in this paper demonstrate the capabilities of dynamic model in improving the sensitivity of the forward kinematic model, resulting in a more even distribution of joint torques and in minimizing the amount of slip between wheels. I. INTRODUCTION A mobile robot with four Powered Caster Wheels (PCW) (Fig. 1) is developed in Singapore Institute of Manufacturing Technology (SIMTech). Mobile robots with Powered Caster Wheels (or off-centered orientable wheels ) have omnidirectional motion capability [1] [] i.e. have three controllable degrees-of-freedom in the planar case. Omnidirectional motion capability simplifies the trajectory planning of the mobile base: all degrees-of-freedom in the planar workspace are controllable simultaneously. Having full maneuverability, the applicability of the mobile base as a service robot in human and unstructured environment is improved. The kinematic and dynamic modelling of PCW-based mobile robots are well studied in the literature [] [6]. Optimal design and analysis of this kind of mobile base are studied in [7] [8]. inematic and dynamic control strategies were formulated and implemented on the omnidirectional base. Various methods were investigated in order to obtain accurate models of the kinematic and dynamic characteristics, therefore achieving better motion control performance. This paper presents the formulation of kinematic and dynamic models of the mobile base and analyzes two forward kinematic models which result in different sensitivities of the base position errors to joint position errors. Using experimental results carried out on the actual robot, the effectiveness of dynamic compensation is demonstrated by achieving a more even joint torque distribution and minimized amount of slip between wheels. These results improve wheel-floor traction, hence decrease the chances of wheel-floor slippage. This in turn produces a more accurate odometry of the base. The strategies described were implemented on the omnidirectional mobile robot and real time experimental results are presented in this paper. Fig. 1. Fig.. An omnidirectional mobile base with 4 Powered Caster Wheels. t Ci r C N rρ i y Ci C i b i y B v p Ci φ 1 φ i ω h i β i β 1 B b x Ci φ N x B Frame assignment and parameter definition of the mobile base. II. INEMATICS The frame assignment and parameter definition of a mobile base with N PCWs are shown in Fig.. Frame B is defined as the frame attached to the center of the mobile base and moves with the base. The location of the wheel with respect to Frame B is defined by vector h i, of length h i, forming an angle β i with the x axis of Frame B. The contact point between the wheel i and the ground is defined as point C i and its position with respect to Frame B is defined as p Ci. A Frame C i is defined with its origin at C 1-78-9177-/5/$./ 5 IEEE 567
point C i, with axis y Ci defined along the translational motion of the wheel and the axis x Ci perpendicular to y Ci so that it results in axis z Ci pointing vertically upwards. Every powered caster wheel is capable of steering and driving motions. The steer angle for wheel number i is defined as φ i and the driving angle as ρ i. The wheel has an offset of length b and radius r. The kinematics of the mobile base can be derived by equating the velocities of the wheel-floor contact points C i as generated by the task space velocity at the center of the base (ẋ =[v,ω] T ) and those generated by the joint space velocities ( q = [ φ 1, ρ 1, φ, ρ,... φ N, ρ N ] T ). When expressed in Frame B, it is described by the following equation: B v+ B ω B p Ci = b φ B i x Ci + r ρ B i y Ci (1) where B x Ci and B y Ci (expressed in Frame B) are [ ] B sin(βi φ x Ci = i ) cos(β i φ i ) [ ] () B cos(βi φ y Ci = i ) sin(β i φ i ) Vector B t Ci is obtained by rotating p Ci by 9 o keeping the same magnitude. The result of the cross product B ω B p Ci is simplified as: ω B t Ci, where ω is the scalar value of the rotation of the base around the vertical z axis (Fig. ). Equation (1) is rearranged into: [ ] [ ] B I t B v Ci = [ ] [ ] b ω B x Ci r B φ y i Ci ρ i () For convenience, () can be expressed with respect to individual Frame C i to reveal the individual contribution of the steer and drive joint motions of each wheel i. This is done by pre-multiplying both sides of the equation with rotation matrix C R B =[ B x B Ci y Ci ] T. The resulting equation of motion for the mobile base, taking into account all the available joints, is expressed as A ẋ = B q (4) where B x T C1 B x T C1 B t C1 B yc1 T B yc1 T B t C1 B x T C B x T C B t C A = B yc T B yc T B t C (5).. B x T C B N x T C B N t CN B yc T B N yc T B N t CN b 1... r 1... B =.... (6)... b N... r N The odd and even rows of (4)-(6) describe the relationship of the velocities at the center of the base with the velocities generated by the steer joint (φ i ) and by the drive joint (ρ i )of wheel i respectively. A. Inverse inematics Equation (4) indicates the relationship between the task space velocities and joint space velocities of the robot. Since B is a diagonal matrix, it is obvious that the inverse kinematics of the robot always exists as: q = B 1 A ẋ (7) The same expression for inverse kinematics can also be derived by modelling each wheel as a serial manipulator with common end-effector point. Each wheel is modelled as serial R P R manipulator with three joints. Expressing the Jacobian in the end-effector frame removes the unobservable and uncontrollable twist (virtual rotation axis at every contact point) joint variable from the inverse kinematic expression for [ φ i, ρ i ] T. The details can be found in [] [4]. B. Forward inematics We have just shown that the inverse kinematics of the mobile base always exists. However, the forward kinematics is not as straightforward. This is due to the fact that the mobile base is a parallel mechanism in topology and the number of joints is greater than the dimension of task space in the system. In our mobile base, every wheel module can be considered as a branch of the whole system and there are more joints (8 joints) than the degrees-of-freedom ( DOFs planar) of the robot. Due to the redundant actuation, the forward kinematics of this type of robot is not straightforward and different optimization schemes have been proposed [9]. In practice, the presence of uncertainties such as sensor noise and parameter error requires statistical methods to compensate for them [1]. The drawback of these methods is that they require either extensive computations in the control process or extra assumptions on the mechanism. In this paper, we will derive two simple forward kinematic models based only on the motion constraints of the robot. Simulation and real experiments in later sections will show the performance of these two simple solutions. 1) Forward inematic Model 1: From the inverse kinematics (7), the forward kinematics can be derived by computing the left pseudo-inverse of matrix B 1 A as ẋ =(B 1 A) LP I q (8) where (B 1 A) LP I =((B 1 A) T (B 1 A)) 1 (B 1 A) T Equation (8) leads to an ẋ solution which minimizes the joint velocity differences between the measured and the consistent set of joint velocities associated with the corresponding robot velocity, in a least-squares manner [11]. ) Forward inematic Model : From (4), the forward kinematics can also be derived by computing the left pseudoinverse of matrix A as: ẋ = A LP I B q (9) where A LP I =(A T A) 1 A T Note that the right hand side of (4) represents the contact point (C i ) velocities generated by the driving and steering 568
actuators on each wheel module. This model will lead to an ẋ which minimizes the differences between the measured contact point velocities and the ideal contact point velocities based on the corresponding robot velocity, in a least-squares way [11]. III. DYNAMICS A. Dynamic Modelling In most of the literature on the study of dynamics of wheeled mobile robots (WMRs), the equations of motion are derived using either the Lagrangian formalism [1] [1] [] or the Newton-Euler formalism [14] [16]. The concept of orthogonal complement was used to develop the dynamics of a DOF WMR [17]. The same concept has been extended to derive the dynamics of an omnidirectional WMR with PCWs [6]. Some other researchers have adopted ane s approach (also termed Lagrange s form of d Alembert s principle) to model the dynamics of WMRs [6] [18]. This approach is shown to offer several advantages over the Newton-Euler and Lagrangian approaches [18]. The Operational Space Formulation [19] is another commonly used method for modelling the dynamics of robot systems. The Augmented Object Model [], an extension of the Operational Space Formulation, is applied to the dynamic modelling of multiple robot systems cooperating to manipulate a common end-effector. The Augmented Object Modelling method is used here to derive the dynamics equation of the mobile base. The mobile base is considered as a cooperative manipulator system. Every wheel module is modelled as a manipulator and the center of the mobile base is the common operational point of the manipulators. The Lagrangian dynamic model in joint space is given as Γ=M q + b + g (1) where Γ is the joint torques vectors, M, b and g are the inertia, Coriolis and centrifugal forces and gravity of the rigid bodies in the system. Similarly, the operational space dynamic model can be derived as F =Λẍ + µ + p (11) where F, µ, p, Λẍ are the resultant force, centrifugal and Coriolis force, gravity and inertial acceleration forces in operational space respectively. The mobile base is assumed to always move in a planar space, so gravity component is ignored in following discussion. The Augmented Object Model states that the dynamics of the composite system is found by summing the operational space dynamics of all the chains and the load at the operational point []. So the augmented inertia Λ, Coriolis and centrifugal force µ are expressed as Λ = Λ i (1) The dynamic formulation of the augmented system in operational space is then given as F =Λ ẍ + µ (14) More details on deriving the dynamic model can be found in [1] and [11]. B. Torque Distribution Two torque distribution schemes can be derived from the two forward kinematic models presented in Section II based on the force/velocity duality. 1) Torque Distribution Scheme 1: From (8), the joint torques Γ can be computed from the operational force F: Γ=(B 1 A) T LP I F (15) By using this model, the joint torque differences are minimized in a least-squares manner. ) Torque Distribution Scheme : Similarly, another torque distribution model can be derived from (9): Γ=B T A T LP I F (16) By using this model, we minimize, in a least-squares manner, the interaction forces between different contact points. The operational space force F is specified by the inverse dynamics of the augmented system given in (14). The control loop can then be closed by choosing one of the torque distribution schemes presented above. IV. SIMULATION AND EXPERIMENTAL RESULTS In this section, the effectiveness of the proposed dynamic control scheme, in terms of its capabilities to evenly distribute joint torques and minimize slippage, are demonstrated. The mobile base (Fig. 1) was commanded to move to and fro in the y direction for meters in 1 seconds in all experiments (Fig. ). For the purpose of comparison, three different control strategies were used in the experiments. The three control strategies are kinematic control ( model) dynamic control with torque distribution scheme 1 ( model) dynamic control with torque distribution scheme ( model) y (m) 1.8 1.6 1.4 1. 1.8.6.4. 1 4 5 6 7 8 µ = µ i (1) Fig.. Desired trajectory of the experiments 569
A. Effects of Dynamic Compensation 1) Tracking Error: Fig. 4 shows the position tracking error of the three models and Fig. 5, expanded view of Fig. 4, shows the position tracking error when the robot is moving in straight line without steering. The following can be observed: In the direction of motion (y), errors for kinematic model are always more than that of the dynamic model (Fig. 4). This is expected as the dynamic model can compensate for the dynamic coupling effects. In the non-motion direction (x), it is interesting to note that the performance of the kinematic model is comparable to that of the dynamic model (See tracking error in x direction in Fig. 5) except when the wheels are flipping and the dynamic coupling is amplified (See tracking error in x direction in Fig. 4). When the wheels change their directions (i.e. when the robot changes its direction by 18 o ), the kinematic model consistently performs worse compared to the dynamic model (peaks in Fig. 4 in the non-motion direction x and θ). This is because the flipping of the wheels amplifies the dynamic coupling effects. The magnitude of errors for the dynamic model is about the same in all directions. This is expected and shows the effectiveness of compensating for the coupling with the dynamic model. x axis (m) y axis (m) theta (deg).1.5 tracking error.5 and.1 5 1 15 5 5 4..15.15. 5 1 15 5 5 4.5 1. 1. and and y axis (m) x axis (m) theta (deg) (a) 4 x 1 4 tracking error 4 6 6.5 7 7.5 8 8.5 9 9.5 1..5 6 6.5 7 7.5 8 8.5 9 9.5 1.1.1. 6 6.5 7 7.5 8 8.5 9 9.5 1 Fig. 5. (b) (c) joint torques (Nm) and Expanded view of Fig. 4 between the two vertical lines. 5 1 15 5 5 4 5 1 15 5 5 4 5 1 15 5 5 4 Fig. 6. Comparison of joint torques (driving) between the three control models (4 curves in each graph, one for each wheel). Drive.5 5 6 1 15 5 5 4 Fig. 4. models. Comparison of position tracking error between the three control ) Torque Distribution: Fig. 6 and 7 show the controller outputs i.e. driving joint torques and steering joint torques under the three control schemes. Both Fig. 6 and 7 show that the torques were much more even across all 8 actuators in the dynamic control (Fig. 6 and 7 middle and bottom) than the kinematic control (Fig. 6 and 7 top). The results from model were also found to be marginally better than the model, although the difference was not as much as that between kinematic and dynamic models. By distributing joint torque evenly, the mobile base increases the overall wheelfloor traction, therefore reducing the chance of wheel-floor slippage. B. Comparison of and The effects of dynamic models and are further investigated model by simulation and real experiments in the following sub-sections. 57
(a) Steer xdot error(m/s).5 sensitivity to disturbance 5 1 15 5 5 4.5 1 4 5 6 7 8 9 1 (b) joint torques (Nm) 5 1 15 5 5 4 ydot error (m/s).5.5 1 4 5 6 7 8 9 1 (c) 5 1 15 5 5 4 thetadot error (deg/s) 4 4 1 4 5 6 7 8 9 1 Fig. 7. Comparison of joint torques (steering) between the three control models (4 curves in each graph, one for each wheel). Fig. 8. Comparison of sensitivity to disturbance between and models. 1) Sensitivity to Disturbance: An interesting difference between the two models is observed in simulation. To demonstrate the sensitivity of the two models to disturbance, a desired base trajectory (ẋ des ) was generated and the desired joint trajectory computed by inverse kinematics. A random disturbance signal ( q) was introduced to the joint trajectory which is then used to compute the base trajectories (ẋ com ) using the two forward kinematic models ( and ). Fig. 8 (a)-(c) show the resulting base motion errors (ẋ des ẋ com ) in x, y and θ respectively. The resulting error in the base motion computed by model is larger than that of model. This is an interesting result which implies that model is provides a more sensitive feedback to the tracking error of the base motion. This is especially so when the errors are computed only from joint position sensors and there is no external sensor for measuring the actual base position. More extensive experimental results have consistently indicated that the model is more sensitive to error or disturbance in joint position. ) Slip Minimization: In this experiment, motion produced by the torque command vector computed using dynamic models and were compared. Fig. 9 shows the measured driving joint velocities ( ρ) of the wheels produced by the resulting torque commands. More spikes in joint velocities are detected in the motion (when the robot is moving in straight line) resulting from model than. These spikes show that not all the wheels are producing the desired motion. Because this is a parallel system, all wheels should ideally produce identical velocity vector at the common operational point (equal to the desired velocity). When this is not the case, the close-chain kinematics no longer holds and slippage occurs between the wheel contacts and the ground. The slippage is due to the fact that the sum of the internal forces produced by the wheels on the mobile base chassis are not zero. When a rigid body is subjected to multiple frictional point contacts (which is the case of our mobile base), the Moore-Penrose Generalized Inverse (or the pseudo-inverse) of the mapping between the operational point and the contact points yields a solution vector which lies in the null-space of the internal force field [1], []. model yields a solution that minimizes the internal forces at the base chassis, thereby slippage is reduced. V. CONCLUSION inematic and dynamic models of a PCW-driven omnidirectional mobile robot are presented. Two forward kinematic models which result in different sensitivities of base position error to joint position error are evaluated. All experiments were carried out on our real omnidirectional mobile base. The experimental results show that, given a desired robot motion, the dynamic control has the capability of distributing joint torque more evenly, as compared to kinematic control. Slippage is also minimized by choosing a solution which minimizes the differences between the measured wheel contact velocities (with the ground) from the desired. These merits of torque distribution and slip minimization enhance the accuracy of mobile base odometry, as they reduce chances of overall wheel-floor slippage. ACNOWLEDGMENT The support of a collaborative research project grant from National University of Singapore and Singapore Institute of Manufacturing Technology (SIMTech) is gratefully acknowledged. We would like to thank Professor Oussama hatib (Stanford University) and Dr Lim Ser Yong (SIMTech) for the discussion and advice and Lim Tao Ming (SIMTech) for all the software support. 571
qdot (rad/s) 15 1 5 5 1 1 4 5 6 7 8 qdot (rad/s) 8 6 4 4 slip is detected 6 1 4 5 6 7 8 Fig. 9. Comparison of slip minimization between the two dynamic control models. In contrast to model, slip was detected in the model. (4 curves in each graph, one for each wheel). [1] B. d Andrwa Novel, G. Bastin, and G. Campion, Modelling and control of non-holonomic wheeled mobile robots, Proc. of IEEE Intl. Conf. on Robotics and Automation, vol., pp. 11 115, 1991. [14] S.-H. Yu and J. J. Moskwa, A global approach to vehicle control: Coordination of four wheel steering and wheel torques, ASME J. of Dyn. Sys., Meas. and Control, vol. 116, no. 4, 1994. [15] N. Matsumoto and M. Tomizuka, Vehicle lateral velocity and yaw rate control with two independent control inputs, ASME J. of Dyn. Sys., Meas. and Control, vol. 114, 199. [16] M. Huang, Dynamic modelling and simulation of an agv (concic ii), Master Thesis, 1991, concordia University, Montreal, Canada. [17] S. umar and J. Angeles, inematics and dynamics of a three-wheeled -dof agv, Proc. of IEEE Intl. Conf. on Robotics and Automation, vol., pp. 157 1577, 1989. [18]. Thanjavur and R. Rajagopalan, Ease of dynamic modelling of wheeled mobile robots (wmrs) using kane s approach, Proc. of IEEE Intl. Conf. on Robotics and Automation, vol. 4, pp. 96 91, 1997. [19] O. hatib, A unified approach for motion and force control of robot manipulators: The operational space formulation, IEEE J. Robotics and Automation, vol. RA-, no. 1, pp. 4 5, 1987. [].-S. Chang, R. Holmberg, and O. hatib, The augmented object model: Cooperative manipulation and parallel mechanism dynamics, Proc. IEEE Intl. Conf. Robotics and Automation, vol. 1, pp. 47 475,. [1] V. R. umar and. J. Waldron, Force distribution in closed kinematic chains, Journal of Robotics and Automation, vol. 4, no. 6, pp. 657 664, December 1988. [] M. A. Nahon and J. Angeles, Force optimization in redundantlyactuated closed kinematic chains, Proc. IEEE Intl. Conf. Robotics and Automation, vol., pp. 951 956, 1989. REFERENCES [1] R. Holmberg and O. hatib, Development and control of a holonomic mobile robot for mobile manipulation tasks, International Journal of Robotics Research, vol. 19, no. 11, pp. 166 174,. [] G. Campion, G. Bastin, and B. D Andrea-novel, Structural properties and clasification of kinematic and dynamic models of wheeled mobile robots, IEEE Transactions on Robotics and Automation, vol. 1, no. 1, pp. 47 6, February 1996. [] M. T. Zaw, D. N. Oetomo, M. H. Ang. Jr., and T.. Ng, inematics and dynamics of an omnidirectional mobile platform with powered caster wheels, In International Symposium on Dynamics and Control, 15- September, Hanoi, Vietnam,. [4] R. Holmberg and O. hatib, Development and control of a holonomic mobile robot for mobile manipulation tasks, International Journal of Robotics Research, vol. 19, no. 11, pp. 166 174,. [5] B. Yi and W. im, The kinematics for redundantly actuated omnidirectional mobile robots, Journal of Robotic Systems, vol. 1, no. 6, pp. 55 67,. [6] J. H. Chung, B.-J. Yi, W.. im, and H. Lee, The dynamic modeling and analysis for an omnidirectional mobile robot with three caster wheels, Proc. IEEE Intl. Conf. Robotics and Automation, vol. 1, pp. 51 57,. [7] T. Park, J. Lee, B. Yi, W. im, B. You, and S. Oh, Optimal design and actuator sizing of redundantly actuated omni-directional mobile robots, Proc. IEEE Int l. Conf. on Robotics and Automation, vol. 1, pp. 7 77,. [8] S. im and H. im, Isotropy analysis of caster wheeled omnidirectional mobile robot, Proc. IEEE Int l. Conf. on Robotics and Automation, vol., pp. 9 98, 4. [9] Y.. Yiu and Z. X. Li, Optimal forward kinematics map for a parallel manipulator with sensor redundancy, Proc. IEEE Intl. Symposium on Computational Intelligence in Robotics and Automation, vol. 1, pp. 54 59,. [1] F. Marquet, F. Pierrot, and O. Company, A statistical approach for the computation of the forward kinematic model of redundantly actuated mechanisms, Proc. IEEE Intl. Conf. on Intelligent Robots and Systems, vol., pp. 558 56,. [11] R. Holmberg, Design and development of powered-caseter holonomic mobile robots, PhD. Dissertation, August, stanford University. [1] Y. Zhao and L. BeMent, inematics, dynamics and control of wheeled mobile robots, Proc. of IEEE Intl. Conf. on Robotics and Automation, vol. 1, pp. 91 96, 199. 57