by Means of Haptic Feedback. Nicola Diolaiti .... It is important to underline that the proposed force-feedback algorithms are not ... wheels whose velocity difference generates the steering mo- tion. A third wheel ..... tional units shown in Fig.
Obstacle Avoidance for Teleoperated Mobile Robots by Means of Haptic Feedback Nicola Diolaiti, Claudio Melchiorri DEIS - Dept. of Electronics, Computer Science and Systems University of Bologna - Italy Email: {ndiolaiti, cmelchiorri}@deis.unibo.it Abstract— The development of remotely operated devices is historically motivated by the need to carry out complex operations in potentially hazardous remote environments. The class of telemanipulation systems includes also mobile robots that can be remotely operated to perform particular tasks. As an example, the inspection of underwater structures and the removal of mines are performed by mobile platforms controlled by a remote operator, which generally takes advantage only of the visual feedback provided by vision systems. In this sense, the nature and completeness of the data provided to the operator about the state of the remote system are of crucial importance for proper task execution, and it is generally accepted that a more efficient achievement of the task can be obtained by increasing the number of data feedback and by using proper MMI. In this paper, the use of a haptic interface is proposed in order to increase the user’s perception of the workspace of the mobile robot. In particular, a virtual interaction force is computed on the basis of obstacles surrounding the mobile vehicle in order to prevent dangerous contacts, so that navigation tasks can be carried out with generally better performances.
I. I NTRODUCTION Teleoperated mobile robots are widely used in order to carry out complex tasks in hazardous environments: well known examples are e.g. the inspection of underwater structures [1], demining operations [2], or cleaning nuclear plants [3]. In this type of apparatus, often the remote operator can take advantage only of visual information about the environment and in the majority of the cases they are not sufficient to carry out complex tasks because of the limited visual fields of cameras. Therefore, besides the possibility of errors and failure of the task, remote teleoperations turn out to be tiring activities requiring a specific training to the human operator. In [4] it is discussed how more sophisticated man-machine interfaces can improve the performances of the overall system by augmenting the number and quality of data feedback from the remote environment. In particular, a noticeable reduction of operator’s stress and of task errors can be achieved by means of a haptic device that allow the operator to perceive forces related to obstacles surrounding the mobile robot. Another application of this concept can be found in [5], where a force sensor is mounted on a mobile vehicle in order to measure the contact force with objects that have to be shifted from a place to another. By means of a suitable MMI, the user can perceive the measured force and therefore he can detect is the object is blocked by an obstacle. The availability of a suitable haptic interface could therefore represent a significant improvement
in the usability of a teleoperated mobile robot, and could make this kind of system suitable for a wide range of operations, i.e. the safe transportation of heavy payloads inside industrial plants. In [6], the distance from obstacles, measured by a laser scanner mounted on a mobile robot, is used to compute a repulsive force that is rendered to the human operator by means of a haptic interface. The haptic device is also used to control the robot motion. In addition, authors present some experimental results, confirming that the augmented perception of the environment surrounding the vehicle reduces the number of collisions with obstacles. In this paper, the problem of safely controlling a remote mobile platform is addressed. Several important aspects are considered: the nonholonomic constraint of the mobile robot, the need to detect by means of low-cost sensors the presence of obstacles, the stability of the overall system, the possible presence of communication time delays. For these reasons, passivity is considered a fundamental aspect in the proposed control strategy. In particular, an IPC (Intrinsically Passive Control) scheme is introduced in order to provide passivity also during interaction with unknown environments [7], [8]. The structure of this paper is the following. In Sec. II the teleoperation system is briefly described and some details on the map-building algorithm are illustrated. In Sec. III the mobile vehicle is modelled as a virtual mass subject to forces exerted by the operator and by the environment. These interaction forces are described in details in Sec. IV, where also a model of the complete system is discussed. Simulations and initial experimental tests are presented in Sec. V, while Sec. VI concludes with final remarks. II. OVERVIEW OF THE S YSTEM The teleoperation system considered in this paper is schematically illustrated in Fig. 1. Data acquired by proper sensors (i.e. sonars) mounted on a Pioneer mobile robot are processed in order to build a local map of the surrounding obstacles. On the basis of this map and of the kinematic status of the vehicle, a virtual interaction force Fe , emulating a physical contact by means of a virtual (repulsive) spring Ke and a virtual damper be , is computed as shown in Fig. 2. The virtual interaction force is sent, by means of a local network using the UDP protocol, to a PHANToM haptic interface in order to render to the operator the feeling that the vehicle
LAN
Haptic interface
Haptic control loop
Map-Building and Robot Control
Mobile robot
Fig. 1. Overview of the teleoperation system: the virtual interaction force computed on the basis of the local map is sent to the haptic interface, whose position is used to compute motion commands sent to the mobile platform.
is close to an obstacle. Conversely, the human operator, by means of the haptic device, generates velocity set-points that are transmitted to the mobile robot controller. It is important to underline that the proposed force-feedback algorithms are not dependent on the described experimental setup, that has to be designed on the basis of specific application requirements in terms of cost and performances. Indeed, our experimental setup is quite modular and each module, i.e. sensor used for the map-building, can be easily changed to meet different cost/performances trade-offs. obstacle
Virtual Interaction Force Fe PSfrag replacements
Ke
s(0) := [0, 0, 0, 1]T be
Mobile Robot
Fig. 2.
where sO and sF quantify, respectively, the belief that the cell is occupied or free, while sU quantifies the belief that the cell is unknown, and quantifies the lack of more detailed information. Finally, s∅ represents the amount of contradictory information accumulated from sonars; indeed, when e.g. a previously occupied cell becomes free, a great amount of contradiction is generated by the combination of the previously accumulated information with new measures coming from sensors, and this fact can be used to quickly detect changes in the environment [10]. Note that the sum of the elements of the status vector has to be 1. Initially, all cells are unknown and their status is:
Virtual interaction with obstacles.
A. Map-building algorithm Information about the environment surrounding the mobile robot is acquired by means of 16 ultrasonic sensors mounted on the Pioneer platform. These sensors, even if widely used in mobile robotics, are affected by several drawbacks. Indeed, the measure of the position of an obstacle is obtained with a poor angular resolution (i.e. about 25 degrees) and multiple reflections of the acoustic wave can occur so that a distance greater than the real one is measured by the sensor. Moreover, the speed of sound limits the maximum sampling rate to about 50 ms. According to these considerations, data provided by each sensor are filtered and then fused into a map of the environment surrounding the vehicle. In particular, the map is represented by a grid of cells, that can be either empty or occupied by an obstacle. In order to represent the occupancy status of each cell, an algorithm based on the Transferable Belief Model [9] has been adopted because it is faster than standard probabilistic methods in detecting changes in the environment. A vector s of basic belief masses is associated to each cell of the grid: s := [s∅ , sO , sF , sU ]T
(1)
Whenever new measures are available from sonars, they are filtered in order to reduce the influence of the poor angular resolution and of multiple reflections and then they are fused into the gridmap by means of the Dempster’s rule of combination [9]. By means of the so called pignistic transformation it is possible to compute the occupancy probability pO of the cell: sU ´ 1 ³ sO + (2) pO = 1 − s∅ 2 Finally, only cells whose occupancy probability is greater than a fixed threshold value are considered occupied and generate a virtual interaction force on the mobile robot. III. C ONTROL S TRATEGY The Pioneer mobile platform used in this paper belongs to the class of two-wheeled robots, because it has two actuated wheels whose velocity difference generates the steering motion. A third wheel, called castor, is not actuated and is used to provide stability to the vehicle. Therefore, the kinematic model is expressed by: # x˙ v cos θv 0 " v(t) y˙ sin θ 0 (3) v v = ω(t) 0 1 θ˙v
where [xv , yv , θv ]T represents the position and the orientation of the vehicle with respect to a fixed reference frame, and [v(t), ω(t)]T represents the translational and rotational velocities. In the hypothesis that the two actuated wheels are constrained to roll without slip over a horizontal plane, feasible trajectory for the mobile robots have to be tangent with
−Fok
LAN
Fo
Interaction with operator
FE
Virtual Interaction FE
Map Building
x˙ o Sonars
Haptic Interface
Virtual Mass
Controller
m
PSfrag replacements Control of the interaction
Fig. 3.
x˙
Conversion of velocities
Mobile Robot
[vd , ωd ]T
Virtual mass Real robot m
Scheme of the proposed control strategy
its translation axis. Indeed, the constraint of rolling without vr slipping is nonholonomic and implies that the translational x˙ x˙ vt velocity v(t) of the mobile robot is always orthogonal to the axis of the actuated wheels, without reducing the set of θv possible configurations of the vehicle, which is
