Dextrous Manipulation with Rolling

[Rolling with Soft Fingertips] [Rolling with Sensor Feedback]

Manipulation with rolling adds complexity as well as flexibility to a dextrous hand. Rolling can increase the reachable workspace of the object and allow tactile sensors to determine the shape of an object. However, rolling can be a problem in that manipulation makes use of a variety of sensors both for event detection and for control. Our approach is to develop a data manager that combines model- and sensor-based information.

[Picture of rolling manipulation] Model-based information uses a model of the object in order to determine the kinematics of rolling. When the shape of both the object and the fingertips are known, the motion of the fingers can be planned to give the desired motion of the object and/or the motion of the contact points, depending on the degrees of freedom of the system. The figure on the right shows several frames (created in a Matlab simulation) where the green object moves through 3 positions while the fingertips roll over the surface. However, if the fingertips are not completely rigid, the effects of unmodeled information must be considered as explained in the following section.

Rolling with Soft Fingertips

When using soft, instrumented fingertips, the kinematic models developed for rigid-body rolling may not be accurate and a soft rolling model may be needed. Local deformation of the fingertips will likely cause the contact points to diverge from their predicted values.

The following figure shows the results obtained when rolling with several soft materials. Interestingly, the effective rolling speed may be increased, decreased, or essentially unchanged compared to the rigid body case, depending on the fingertip construction.

Details on rolling with soft fingertips can be found in:

[Chang and Cutkosky 1995], "Rolling with Deformable Fingertips", 1995 IEE/RSJ Intl. Conference on Intelligent Robots and Systems
Dean Chang's PhD thesis

Rolling with Sensor Feedback in a Phase-Based Framework

Rolling Phases

In the Phase/Event/Transition Framework, rolling is accomplished as a coordinated motion phase in which both the motion of the object and the trajectories of the contact points over the surface of the object can be controlled. The rolling phases can be split into two main categories, explicit and implicit.

[Type of rolling phases] Explicit rolling is a phase where we directly specify the trajectories of the object and the contact points on the object or fingertip surface. Contact point trajectories are related to the object motion through the surface geometry of the the object and the fingertips [Montana 1988].

In Implicit rolling we specify the trajectory of the grasped object only; the trajectories of the contact points are automatically chosen to minimize a task-related objective function. Examples would be keeping grasp forces within the friction cones at the contact points (in order to maintain a stable grasp) or minimizing the possibility of reaching a joint limit. The current focus of our work is on explicit rolling.

It should be noted that the ability of a dextrous hand to manipulate in these modes depends on the degrees of freedom of the system. A planar two-fingered hand that has 3 dof fingers has five degrees of freedom when grasping an object, so the contact positions, as well as the object position and orientation, may be specified. However, a planar two-fingered hand with 2 dof fingers only has 3 degrees of freedom when grasping an object. When object position and orientation are specified, we cannot control the motion of the contact points. Thus, such a system would require implicit rolling with no objective function.

Rolling an Unknown Object using Sensor Feedback

An interesting task made possible by rolling is the manipulation of an unknown or incompletely known object. Using a rolling data manager (see figure at top) to combine various sensor- and model-based information, we may obtain a detailed model of the local geometry of the object while manipulating it in a controlled manner.

In particular, a tactile array sensor may be used to get an estimate of local surface geometry properties, such as radius of curvature. The figure above gives an example of a shape recognition technique. Although it is possible to accomplish rolling of an unknown object without such estimates [Maekawa, Tanie, and Komoriya 1995], the ability to accurately control the motion of the object and the contact points is greatly improved if this information is made available.

The following simulation shows how sensed radius of curvature can be used. The object being manipulated is an ellipse, which has a continuously varying radius of curvature at all points except at the major and minor axes. This dextrous hand, with flat fingertips and two 3 dof fingers, has been commanded to move the unknown object up 5 cm and rotate it 30 degrees while moving each contact point 2 cm over the object surface. It starts out holding the object, but the system has no initial knowledge of the shape except the position of the arbitrarily assigned center of the object and the local curvature.

See an animated simulation or some overlaid frames.

Application of Rolling with Sensor Feedback in the Phase-Based Framework

[Phase/Event/Transition Structure for Rolling] The figure on the right shows the phase/event structure for a rolling task and a close-up of the rolling phases.

The incorporation of rolling with tactile array sensor information into the phase-based framework can be accomplished as follows. The inputs to a rolling phase within a manipulation task will specify the desired motion of the object and the contact points. The task is divided into increments over which the object may be assumed to have constant radius of curvature. The size of these increments can be predetermined by the task or it can depend on the sensed curvature.

The rolling phase will have two possible done events: the completion of an incremental roll or the completion of the entire task. The completion of an incremental roll will trigger another rolling phase that will read the sensor data, plan the next move, then execute it. The planning will use the latest data about the curvature of the object and the positions of the contact points to plan a trajectory for the remainder of that phase. The completion of the entire task will leave that rolling phase completely.

The acquisition of data from the tactile array is somewhat time consuming, so we make special provisions within the coordinated rolling phase class to accommodate it. As a subclass of the action phase, a rolling phase has an explicit start-up segment as shown in the framework description . We can acquire tactile sensor data during the startup segment. If necessary, time-consuming activities such as dynamics calculations can be suspended during this time.

Back to the Dextrous Manipulation Lab home page or projects page.

Dean Chang and Allison Okamura
November 1995



						Dextrous Manipulation with Rolling [Rolling with Soft Fingertips] [Rolling with Sensor Feedback] Manipulation with rolling adds complexity as well as flexibility to a dextrous hand. Rolling can increase the reachable workspace of the object and allow tactile sensors to determine the shape of an object. However, rolling can be a problem in that manipulation makes use of a variety of sensors both for event detection and for control. Our approach is to develop a data manager that combines model- and sensor-based information. Model-based information uses a model of the object in order to determine the kinematics of rolling. When the shape of both the object and the fingertips are known, the motion of the fingers can be planned to give the desired motion of the object and/or the motion of the contact points, depending on the degrees of freedom of the system. The figure on the right shows several frames (created in a Matlab simulation) where the green object moves through 3 positions while the fingertips roll over the surface. However, if the fingertips are not completely rigid, the effects of unmodeled information must be considered as explained in the following section. Rolling with Soft Fingertips When using soft, instrumented fingertips, the kinematic models developed for rigid-body rolling may not be accurate and a soft rolling model may be needed. Local deformation of the fingertips will likely cause the contact points to diverge from their predicted values. The following figure shows the results obtained when rolling with several soft materials. Interestingly, the effective rolling speed may be increased, decreased, or essentially unchanged compared to the rigid body case, depending on the fingertip construction. Details on rolling with soft fingertips can be found in: [Chang and Cutkosky 1995], "Rolling with Deformable Fingertips", 1995 IEE/RSJ Intl. Conference on Intelligent Robots and Systems Dean Chang's PhD thesis Rolling with Sensor Feedback in a Phase-Based Framework Rolling Phases In the Phase/Event/Transition Framework, rolling is accomplished as a coordinated motion phase in which both the motion of the object and the trajectories of the contact points over the surface of the object can be controlled. The rolling phases can be split into two main categories, explicit and implicit. *Explicit* rolling is a phase where we directly specify the trajectories of the object and the contact points on the object or fingertip surface. Contact point trajectories are related to the object motion through the surface geometry of the the object and the fingertips [Montana 1988]. In *Implicit* rolling we specify the trajectory of the grasped object only; the trajectories of the contact points are automatically chosen to minimize a task-related objective function. Examples would be keeping grasp forces within the friction cones at the contact points (in order to maintain a stable grasp) or minimizing the possibility of reaching a joint limit. The current focus of our work is on explicit rolling. It should be noted that the ability of a dextrous hand to manipulate in these modes depends on the degrees of freedom of the system. A planar two-fingered hand that has 3 dof fingers has five degrees of freedom when grasping an object, so the contact positions, as well as the object position and orientation, may be specified. However, a planar two-fingered hand with 2 dof fingers only has 3 degrees of freedom when grasping an object. When object position and orientation are specified, we cannot control the motion of the contact points. Thus, such a system would require implicit rolling with no objective function. Rolling an Unknown Object using Sensor Feedback An interesting task made possible by rolling is the manipulation of an unknown or incompletely known object. Using a rolling data manager (see figure at top) to combine various sensor- and model-based information, we may obtain a detailed model of the local geometry of the object while manipulating it in a controlled manner. In particular, a tactile array sensor may be used to get an estimate of local surface geometry properties, such as radius of curvature. The figure above gives an example of a shape recognition technique. Although it is possible to accomplish rolling of an unknown object without such estimates [Maekawa, Tanie, and Komoriya 1995], the ability to accurately control the motion of the object and the contact points is greatly improved if this information is made available. The following simulation shows how sensed radius of curvature can be used. The object being manipulated is an ellipse, which has a continuously varying radius of curvature at all points except at the major and minor axes. This dextrous hand, with flat fingertips and two 3 dof fingers, has been commanded to move the unknown object up 5 cm and rotate it 30 degrees while moving each contact point 2 cm over the object surface. It starts out holding the object, but the system has no initial knowledge of the shape except the position of the arbitrarily assigned center of the object and the local curvature. See an animated simulation or some overlaid frames. Application of Rolling with Sensor Feedback in the Phase-Based Framework The figure on the right shows the phase/event structure for a rolling task and a close-up of the rolling phases. The incorporation of rolling with tactile array sensor information into the phase-based framework can be accomplished as follows. The inputs to a rolling phase within a manipulation task will specify the desired motion of the object and the contact points. The task is divided into increments over which the object may be assumed to have constant radius of curvature. The size of these increments can be predetermined by the task or it can depend on the sensed curvature. The rolling phase will have two possible done events: the completion of an incremental roll or the completion of the entire task. The completion of an incremental roll will trigger another rolling phase that will read the sensor data, plan the next move, then execute it. The planning will use the latest data about the curvature of the object and the positions of the contact points to plan a trajectory for the remainder of that phase. The completion of the entire task will leave that rolling phase completely. The acquisition of data from the tactile array is somewhat time consuming, so we make special provisions within the coordinated rolling phase class to accommodate it. As a subclass of the action phase, a rolling phase has an explicit start-up segment as shown in the framework description . We can acquire tactile sensor data during the startup segment. If necessary, time-consuming activities such as dynamics calculations can be suspended during this time. Back to the Dextrous Manipulation Lab home page or projects page. Dean Chang and Allison Okamura November 1995 {sub_heading}