ओओ3/ओओओओओओओ**/ओओ/ओ<ओ38ओ
`Carnegie Mellon University
`ओओ8ओओ3ओओओओओAओओ8ओ
`Research Showcase
`
`+/0'010# $,. ,73.# #/#.!&
`Institute for Software Research
`
`   
`1—1—1995
`
`!&,,) ,$ ,*-10#. !'#+!#
`School of Computer Science
`
`/ओओओ8<ओओ <ओ*<ओओ <
`Vision and Force Driven Sensorimotor Primitives
`Aओओ38ओओ<3"ओ*
`for Robotic Assembly Skills
`ओओओओओओ!
`J. Dan Morrow
`    
`Carnegie Mellon University
`
`ओ<"ओओओओओ<
`Bradley]. Nelson
`    
`Carnegie Mellon University
`
`ओ<<ओओ
`Pradeep Khosla
`     -(&,/)!*1#"1
`Carnegie Mellon University, pkhosla@cmu.edu
`
`#!,**#+"#" '00',+
`Recommended Citation
`,..,3  + #)/,+ .")#4  +" &,/) ."##- '/',+ +" ,.!# .'2#+ #+/,.'*,0,. .'*'0'2#/ $,. , ,0'! //#* )4
`Morrow, J. Dan; Nelson, Bradley].; and Khosla, Pradeep, "Vision and Force Driven Sensorimotor Primitives for Robotic Assembly
`('))/  

    
` -#. 

`
`Skills" (1995). Institutefor Software Research. Paper 574.
`&8-.#-,/'0,.4!*1#"1'/.

`
`http: //repository.cmu.edu/isr/574
`
`
`
`This Conference Proceeding is brought to you for free and open access by the School of Computer Science at Research Showcase. It has been accepted
`6'/ ,+$#.#+!# .,!##"'+% '/ .,1%&0 0, 4,1 $,. $.## +" ,-#+ !!#// 4 0&# !&,,) ,$ ,*-10#. !'#+!# 0 #/#.!& &,3!/# 0 &/ ##+ !!#-0#"
`for inclusion in Institute for Software Research by an authorized administrator of Research Showcase. For more information, please contact research—
`$,. '+!)1/',+ '+ +/0'010# $,. ,73.# #/#.!& 4 + 10&,.'5#" "*'+'/0.0,. ,$ #/#.!& &,3!/# ,. *,.# '+$,.*0',+ -)#/# !,+0!0 .#/#.!&
`showcase@andrew.cmu.edu.
`/&,3!/#+".#3!*1#"1
`
`Page 1 of 8
`Page 1 of 8
`
`ABB Inc.
`ABB Inc.
`ABB Inc.
`EXHIBIT 1013
`EXHIBIT 1005
`
`EXHIBIT 1013
`
`

`

`Vision and Force Driven Sensorimotor Primitives for Robotic Assembly Skills
`
`In the Proceedings of the 1995 IEEE/RSJ Int. Conf. on
`Intelligent Robots and Systems, Pittsburgh, PA
`August 5-9, 1995, Vol. 3, pp. 234-240
`
`J. Daniel Morrow†
`
`Pradeep K . K hosla‡
`
`Bradley J. Nelson†
`†Robotics Ph.D. Program
`‡Dept. of Electrical and Computer Engineering
`Carnegie Mellon University
`Pittsburgh, PA 15213
`
`Abstract
`
`Integrating sensors into robot systems is an important step
`towards increasing the flexibility of robotic manufacturing
`systems. Current sensor integration is largely task-specific
`which hinders flexibility. We are developing a sensorimotor
`command layer that encapsulates useful combinations of
`sensing and action which can be applied to many tasks
`within a domain. The sensorimotor commands provide a
`higher-level in which to terminate task strategy plans,
`which eases the development of sensor-driven robot
`programs. This paper reports on the development of both
`force and vision driven commands which are successfully
`applied to two different connector insertion experiments.
`
`1 Introduction
`
`Creating sensor-based robot programs continues to be a
`formidable challenge. Two contributing
`factors are
`programming difficulty and lack of sensor integration.
`A ddressing these problems simultaneously is important
`because they are coupled: introducing sensors exacerbates
`the programming problem by increasing its complexity. We
`propose the development of a sensorimotor layer which
`bridges the robot and sensor spaces with the task space. In
`this paper, we introduce the ideas behind sensorimotor
`primitive (SMP) development and provide examples of
`both force and vision driven SMP’s. In addition, some of
`these SMP’s are implemented and used to construct sensor-
`based control strategies for executing two different
`connector insertions (D and BNC).
`Our goal is to build a richer set of command primitives
`which effectively integrate sensing into the command set
`for a particular class of tasks (e.g. rigid-body assembly).
`The goal is to provide the task programmer with higher-
`level commands which
`incorporate sensing and are
`relevant to the task domain. The critical benefits of sensor
`integration are 1) hiding low-level details of processing
`sensor information, and 2) embedding generic task domain
`knowledge into the command set. The first goal is achieved
`by creating small, reconfigurable modules which process
`
`Task Space
`
`Low Dimensional
`
`Sensorimotor
`Primitive Space
`
`High Dimensional
`
`action space
`
`sensor space
`Figure 1: Sensorimotor Space
`and use sensor information; this allows us to leverage the
`application of a sensor since it is encapsulated. The second
`goal is achieved when the models used to interpret sensor
`information are applicable to many tasks within a domain.
`Whenever a sensorimotor primitive is developed, some
`knowledge about the task (and about the domain if the
`knowledge is sufficiently general) is encapsulated as well.
`The problem is identifying common models for sensor
`interpretation which apply to a variety of related tasks. To
`the extent that these models are applicable only to very
`similar tasks, the task domains will be exceedingly small
`for which the command set is applicable. The challenge is
`to construct a sensor-integrated command set with enough
`embedded knowledge to reduce the difficulty of the
`programming task while retaining enough generality to
`have wide applicability.
`
`1.1 Related Work
`
`Many researchers [16][19] refer to skill libraries or
`task-achieving behaviors as a source of robust, skill-
`achieving programs. This postpones (but does not remove)
`the very difficult issue of how to synthesize such skill
`libraries. The sensorimotor layer is a direct effort to ease
`the programming of robust, sensor-based skills.
`Page 2 of 8
`
`

`

`Other researchers have suggested robot control
`primitives. Lyons
`[10] has developed a theory of
`computation for sensor-based robotics; his robot schema
`computational element is very similar to our Chimera
`reconfigurable module [17]. Brockett [1] suggests a
`postscript-type programming language for robotics in
`which task strategies can be described independently of a
`particular robot system. Deno et al [3] discuss control
`primitives which are inspired by the hierarchical nature of
`the neuromuscular system, but these do not have a strong
`connection to the task. Paetsch and von Wichert [14] apply
`a set of heuristic behaviors in parallel to perform peg
`insertion with a dextrous hand. Smithers and Malcolm [16]
`suggest behavior-based assembly as an approach in which
`uncertainty is resolved at run-time and not during planning,
`but they do not address the issue of behavior synthesis.
`Most of these approaches to primitives (except [14]) are
`either task-specific or robot-centered. We are building on
`portions of this past work to make a stronger and more
`general connection of sensor-based control primitives to a
`task domain.
`Planning robot motions based on geometric models [8]
`has been pursued as a method of task-level programming
`which reduces the programming burden. A problem with
`this method is that resulting strategies often fail because of
`inevitable errors in the task model used for planning. We
`believe, like Smithers and Malcolm [16], that uncertainty
`should be resolved at run-time, not during planning. Morris
`and Haynes [11] have argued that geometric models do not
`contain enough information about how to perform a task.
`They argue for using the contact constraints of the
`assembly task as key indicators for guiding task strategies.
`Similarly, we base our sensor-driven primitives on task
`constraints.
`Much work in using force feedback has centered on
`detailed contact models [20]. Schimmels and Peshkin [15]
`have synthesized admittance matrices for particular tasks.
`Strip [18] has developed some general methods for peg
`insertion based on contact models. Donald [4] developed
`methods to derive plans based on error detection and
`recovery which are guaranteed to either succeed or
`recognizably fail. Erdmann [5] has investigated task
`information requirements through abstract sensor design.
`Castano and Hutchinson [6] have proposed task-based
`visual servoing in which virtual constraints, based on the
`task, are maintained. Canny and G oldberg [2] have been
`exploring RISC (reduced intricacy in sensing and control)
`robotics in which simple sensing and action elements are
`coupled. Many of these approaches focus on developing
`sensor use strategies for a particular task. We are trying to
`generalize sensor use for a task domain by building sensor-
`driven commands which are based on common task
`constraints in both vision and force.
`
`2 Trajectory Primitives
`
`Trajectory primitives are encapsulations of robot
`trajectory specifications. We have developed
`three
`trajectory primitives which are used in our experiments.
`The movedx primitive applies a cartesian velocity over time
`to achieve the specified cartesian differential motion. The
`ldither (rdither) primitive implements a linear (rotary)
`sinusoidal velocity signal at the specified frequency for the
`specified number of cycles. This is useful during assembly
`operations to locally explore.
`Complex trajectories can be specified by combining
`trajectory primitives. For example, combining sinusoidal
`dithers in orthogonal directions can be used to implement
`an “exploration” of an area; the resulting position patterns
`are called Lissajous figures. In order to densely cover an
`area, the frequency ratio (n>1) between orthogonal dithers
`should be selected as (N+1)/N, where N is the number of
`cycles (of the smaller frequency sine wave) before the
`Lissajous pattern repeats. Figure 2 shows the Lissajous
`figures for two orthogonal dither signals with different
`values of the frequency ratio, n. Note that the positional
`space is well-covered by these patterns. A smaller n (closer
`to 1) provides more dense coverage but requires more
`cycles (and hence longer time) to execute.
`
`1
`
`0.5
`
`1
`
`0.5
`
`-1
`
`-0.5
`
`
`
`00
`
`
`
`00
`
`0.5
`
`1
`
`-1
`
`-0.5
`
`
`
`00
`
`
`
`00
`
`0.5
`
`1
`
`-0.5
`
`-1
`
`-0.5
`
`-1
`
`(b) 5 cycles
`(a) 3 cycles
` n=1.2
`n=1.333
`Figure 2: Lissajous patterns
`
`3 Sensorimotor Primitives
`
`is a parameterized
`sensorimotor primitive
`A
`encapsulation of sensing and action which can be used to
`build task strategies or skills. A skill is a particular
`parameterized solution to a specific task (e.g. peg in hole)
`and is composed of sensorimotor primitives. In order to
`develop sensorimotor primitives, the common element(s)
`which relate tasks in the domain must be identified. For
`assembly tasks, force-driven primitives which provide for
`the acquisition, maintenance, and detection of different
`types of contact constraints are useful. Likewise, vision-
`driven primitives can be used to enforce positioning
`constraints which can be sensed on the image plane. We
`rely on these constraints in both the force and vision spaces
`to provide guidelines for sensor-driven commands.
`Page 3 of 8
`
`

`

`3.1 Vision-Driven Primitives
`
`centering and obstacle avoidance could be implemented
`with such a primitive.
`
`3.2 Force-Driven Primitives
`
`Guarded move. This primitive is the common guarded
`move in which straight-line motion is terminated by
`contact. The contact is detected by a force threshold in the
`direction of motion. The basic constraint model is a
`transition from free space (and complete motion freedom)
`to a point/plane contact where one direction DOF has been
`removed.
`Sticking move. The “stick” primitive involves the
`transition, upon contact, to a “maintenance” velocity which
`will maintain contact with a surface when used with a
`damping controller. In addition, the cartesian position is
`monitored in the direction of the maintenance velocity, and
`if the displacement exceeds a specified threshold, the
`primitive detects this and terminates. This prevents the
`robot from continuing to move when contact has been lost,
`and encapsulates the maintenance of point/plane contact
`with loss detection.
`Accommodation. The sub-matrices of a 6x6 damping
`matrix which provides accommodation control can be
`viewed as sensorimotor primitives. The most common one
`is linear accommodation: complying to linear forces by
`performing translations. A sensorimotor primitive which
`introduces angular accommodation in response to torques
`and forces implements a remote-center-of-compliance [20]
`useful for peg insertion tasks.
`Correlation. A ctive sensing primitives, which use the
`commanded action to process the sensor signal, are
`effective ways of extracting information from biased and
`noisy sensor signals [7]. We have employed a correlation
`technique to detect when the reference command is
`perturbed by the damping controller,
`indicating the
`presence of a motion constraint. The correlation (C) is
`computed with the following equation:
`⎞
`⎛
`f i2π
`g i2π
`⎞
`⎞
`⎛
`⎛
`⎟
`⎜
`--------
`--------
`N
`⎠
`⎠
`⎝
`⎝
`⎠
`⎝
`N
`N
`i 0=
`-----------------------------------------------------------------
`f i2π
`g i2π
`⎛
`⎞
`⎛
`⎞
`--------
`--------
`⎝
`⎠
`⎝
`⎠
`N
`N
`
`(1)
`
`N∑
`
`i 0=
`
`The vision-driven primitives are based on visual
`servoing techniques. A n image-based visual servoing
`approach is used, rather than a position-based approach, to
`avoid calculating the inverse perspective mapping of the
`scene at each sampling period. Thus, we must provide
`reference inputs to our visual servoing system in feature
`coordinates. To do this, desired 3D object positions must be
`mapped into image coordinates using a simple perspective
`projection model of the particular visual sensor. These
`primitives enforce positioning constraints on the task using
`poorly-calibrated camera/robot systems; errors on the
`image plane are used to drive the manipulator. A complete
`description of this visual servoing method can be found in
`[13].
`Vision primitives are used to enforce positioning
`constraints on the image plane which are relevant to the
`task. The effective bridge between task space and robot/
`sensor is constructed by enforcing key task positioning
`constraints which can be sensed on the image plane by
`tracking and controlling a finite number of critical points in
`the task.
`fundamental vision
`Image plane translation. A
`primitive is the resolution of image-plane errors through
`translation
`commands;
`this
`enforces
`“translation”
`constraints in the image plane. Castano and Hutchinson [6]
`have proposed a similar method to perform tasks. We use
`this primitive with a two-camera arrangement to enforce 3
`DOF position constraints. In addition, the primitive is
`written so that individual combinations of axes can be
`controlled. This enables the flexibility needed to de-couple
`translation commands for certain tasks (e.g. grasping along
`a specific approach direction). This primitive was
`implemented and used in the connector insertion strategies.
`Image plane rotation. A nother common positioning
`primitive is to align lines in the image plane. Insertions, for
`example, can be very sensitive to errors in the insertion axis
`alignment. A n edge-detection algorithm can robustly
`extract edges from an image and a primitive can use this
`information along with an approximate task model to align
`edges.
`Fixed point rotation. Rotation about the normal of the
`image plane causes a translation of all points not on the
`optical axis. Therefore, one primitive involves selecting a
`particular point fixed with respect to the end-effector which
`is relevant to the task and maintaining its position (in the
`image) during a rotation.
`Visual grasping. One primitive which can be very
`useful is a vision-guided primitive from a camera mounted
`on the gripper -- so-called “eye-in-hand” primitives. This
`can be used to align the gripper with cross-sections which
`are extracted from binary vision images. A utomatic
`Page 4 of 8
`
`N∑
`
`N∑
`
`i 0=
`
`C
`
`=
`
`For two fully correlated sinusoidal signals, the
`correlation value is π2/8. Because the correlation technique
`is based on phase differences, the normalization is required
`to compensate for magnitude changes in the signals which
`affect the computed value.The correlation value is tested
`against a threshold and an event is triggered when the
`correlation drops below
`the
`threshold. The
`full
`development of this primitive is discussed in [12].
`
`

`

`4 Experimental Results
`
`Our experimental results are based on two connector
`insertions: a 25-pin D-shell connector and a BNC
`connector. Figure 3 shows diagrams of the part of each
`connector held by the gripper. Both connectors were stably
`grasped by our pneumatic, two-fingered gripper; no special
`fingers were constructed.
`
`vis_xz
`
`vis_xyz
`
`grip
`
`movedx
`
`gmove
`
`vis_xyz
`
`ldither
`correlate
`
`stick
`rdither
`
`grip
`
`movedx
`
`Insert
`Transport
`Grasp
`Figure 5: D-connector Insertion Strategy
`
`vis_xz
`
`vis_xyz
`
`grip
`
`movedx
`
`vis_xyz
`
`gmove
`
`ldither
`correlate
`
`stick
`ldither
`
`movedx
`
`stick
`
`movedx
`
`Y
`
`X
`X
`
`Y
`
`X
`
`X
`
`Z
`D-connector
`
`Z
`BNC connector
`
`Figure 3: Connector Diagrams
`The strategies are implemented as finite-state machines
`(FSM). Figure 4 shows an example FSM strategy in task-
`space which accesses the primitives in the sensorimotor
`space. The connector insertion strategies (Figure 5 and
`Figure 6) are shown in the task space with the primitives
`explicitly shown in the FSM. G iven the small scale of the
`contact, we cannot reasonably derive strategies based on
`detailed contact-state analyses. Instead, heuristic strategies
`were developed based on the available command set and
`sensing. The strategies are based on the available command
`primitives (some sensor-driven, some not) and are
`implemented as finite-state machines (FSM). A lthough the
`different connector geometries lead to very different
`strategies, the same command primitives can be used to
`implement these strategies.
`
`D-connector Skill
`
`Task
`Space
`
`grip
`movedx
`Insert
`Transport
`Grasp
`Figure 6: BNC Connector Insertion Strategy
`In Figure 5 and Figure 6, each of the “bubbles” in the
`FSM is a Chimera module implementation of a real-time
`computing process. The vis_ modules implement visual
`servoing primitives; for example, vis_xz implements visual
`servoing along the x and z axes. The grip module operates
`the gripper. The other modules (gmove, stick, movedx,
`ldither, rdither) are described in the force-driven or
`trajectory primitive
`sections. The
`task
`strategy
`implemented by primitives results in a command velocity,
`V cmd, which is perturbed by the accommodation controller
`to permit contact. The perturbed velocity, Vref, is used to
`generate joint setpoints for the robot joint controller. A ll of
`the experimental results are shown as plots of Vref.
`For each connector insertion task (Figure 5 and Figure
`6) the strategy involves three phases: 1) grasp the
`connector, 2) transport to the mating connector, and 3)
`perform the insertion. The grasp and transport steps are
`dominated by vision-feedback; the
`insertion step
`is
`dominated by force feedback. The first step, grasping,
`relies on approximate angular alignment of the connector
`axes (X, Z) with the camera optical axes. Visual setpoints
`are identified in the images and controlled through visual
`feedback. The transport step also involves using visual
`feedback to position the grasped connectors above the
`mating connector for insertion. The insertion step is
`different
`for each
`task because of
`their different
`geometries; however, these two different strategies are
`implemented with the same set of primitives. For the D-
`Page 5 of 8
`
`movedx
`
`r d i t h e r
`
`l d i t h e r
`
`gmove
`
`stick
`
`Σ
`
`Accommodation
`
`Controller
`
`Robot
`
`Force
`Sensor
`
`correlate
`
`Sensorimotor
`Space
`
`Robot/Sensor
`Space
`
`Figure 4: Finite-State Machine Strategy
`
`

`

`connector, the insertion begins with a guarded move to
`make contact with the top of the connector. This is followed
`by a sticking move (to maintain the contact) and a mixture
`of rotational and linear sinusoidal dithering (at different
`frequencies), and correlation monitoring of the linear
`dithering. The dithering introduces enough variation in the
`command to resolve small uncertainties left over from
`initial positioning. The correlation of the commanded and
`force-perturbed reference velocities provides a means to
`reliably detect when the connector has seated. This
`“success-detection” method does not rely on attaining
`absolute position goals, but rather on attaining and
`detecting a motion constraint. The strategy for the BNC
`connector begins similarly, with a guarded move. This is
`followed by a sticking move (to maintain contact) and a
`mixture of two linear dithers to implement a Lissajous
`pattern exploration around the insertion axis. We found this
`to be necessary as the vision primitive was not able to
`reduce positioning errors enough to guarantee insertion
`using a (nominal) straight-line move. A gain, correlation is
`used to detect when the connector “seated.” For the BNC,
`however, there is an additional step: the bayonet shaft stubs
`must be mated. This can be performed with a 180 degree
`rotation and terminated using the stick primitive to detect
`when the connector advances along the insertion axis.
`Finally, a 90 degree rotation about insertion axis locks the
`bayonet connector.
`Figure 8 and Figure 7 show cartesian velocity traces for
`experimental trials for each connector insertion task. The
`three stages of the tasks are labelled in each plot and the
`breakpoints for different stages are sensor-driven. The use
`of both force and vision in performing these tasks
`significantly
`increases
`their robustness. Earlier D-
`connector results using only force feedback [12] required
`more stringent initial positioning requirements and ignored
`
`velocity (m/s; rad/s)
`
`0.02
`
`0.01
`
`0
`
`-0.01
`
`grasp
`
`Vz
`Vx
`
`Vy
`
`the connector grasping phase of the task. Visual feedback
`makes the (final) insertion step very reliable since position
`errors are significantly reduced.
`The grasp stage is essentially the same for both
`connectors. Figure 9 shows a vision-guided grasp of a
`connector with three distinct stages: the approach in X and
`Z, followed by the approach along -Y, followed by the
`depart move along +Y. Continuing to visually servo the X
`and Z directions during the approach along -Y is important
`to compensate for calibration errors; this is clearly shown
`in Figure 9 where the V x and V z are non-zero toward the
`end of the -Y approach move. Figure 9 also clearly shows
`when the part is grasped and the final depart move. The
`transport stage is very similar except that it does not have
`the grasp and depart phases.
`Figure 10 shows a close-up view of the D connector
`insertion results from Figure 7. The key parts of the
`strategy are labelled in the plot. There is a 0.5s delay during
`the grip primitive to allow the fingers time to open. Figure
`11 shows a close-up view of the BNC insertion stage results
`from Figure 8. The key parts of the coax insertion strategy
`are labelled in the plot. In this case the first insertion stage,
`which is detected through correlation, was achieved almost
`immediately. In order to compute a proper correlation
`value, at least 1 full cycle of dithering is completed before
`the correlation primitive generates a valid correlation value
`(hence the dithering will not be terminated for at least one
`cycle). Clearly visible on the plot is the movement along -
`Y towards the end of the first rotation. This is detected by
`the stick primitive and signals the transition to the last
`rotation state which locks the bayonet connector.
`Both of these strategies were easily described with
`sensorimotor primitives and were successful in repeatedly
`performing the tasks in spite of imprecisely calibrated
`
`transport
`
`insert
`ωy/15
`
`Velocity (m/s; rad/s)
`
`0.02
`
`0.015
`
`0.01
`
`0.005
`
`0
`
`-0.005
`
`-0.01
`
`V x
`
`V z
`
`15
`
`20
`
`25
`
`30
`
`35
`
`40
`
`45
`
`time (sec)
`Figure 7: D-connector Insertion Results
`
`grasp
`
`Vy
`
`transport
`
`insert
`ωy/15
`
`10
`
`20
`
`30
`
`time (sec)
`
`40
`
`50
`
`Figure 8: BNC Connector Insertion Results
`
`Page 6 of 8
`
`

`

`sensors and lack of firm fixturing. The most crucial error,
`insertion axis alignment, was usually small from the
`beginning position of the parts (the connector was grasped
`standing, not lying on its side). Nonetheless, the rapid
`motion of the pneumatic gripper fingers sometimes
`introduced some skew
`in the
`insertion axis angle.
`Fortunately, the guarded move and subsequent dithering
`usually re-aligned the axis so that insertion proceeded
`without error. However, the tendency of the connector to
`rotate about the task Z-axis, due to only friction coupling,
`made the correlation detection threshold more difficult to
`set. If the part is firmly fixtured, one expects the correlation
`of command and reference signals to drop sharply when
`motion constraints are encountered (e.g. when seating the
`
`connector). However, our lack of firm fixturing resulted in
`significant motion of the task box during these stages of the
`task.A larger correlation threshold had to be set in order to
`reliably detect the seating stage. More precise fixturing and
`tooling would alleviate these problems considerably, but
`our strategies were able to succeed in spite of them. Besides
`the common guarded move, the stick primitive, which
`encapsulates contact maintenance, and the correlation
`primitive, which detects the difference in commanded and
`reference velocity signals, proved very useful in both tasks.
`The stick command was used differently in the two tasks.
`In the D-connector, it was used to detect an error condition
`if contact was lost; in the BNC, it was used to detect when
`the bayonet “mated” with connector shaft stubs.
`
`Y
`
`X
`
`Z
`
`Velocity (m/s)
`
`0.02
`
`0.015
`
`0.01
`
`0.005
`
`0
`
`-0.005
`
`-0.01
`
`visually servo in
`X and Z only
`
`V x
`
`visually servo in
`X, Y, and Z.
`
`calibration error
`compensation
`
`V z
`
`4
`
`6
`
`8
`
`10
`
`Vy
`
`pick up
`connector
`
`gripper closes
`here
`
`12
`
`14
`
`16
`
`18
`
`20
`
`22
`
`Velocity (m/s)
`
`time (sec)
`
`Figure 9: Vision-Guided Grasp
`
`0.02
`
`0.01
`
`0
`
`-0.01
`
`end of transport phase
`
`dithering
`
`depart move
`
`35
`
`36
`
`37
`
`guarded move
`
`38
`
`contact
`
`39
`
`40
`
`41
`
`42
`
`43
`
`44
`
`time (sec)
`
`gripper opened
`correlation threshold
`tripped
`
`0.02
`
`0.01
`
`0
`
`-0.01
`
`Velocity (m/s; rad/s)
`
`end of transport phase
`dithering
`
`Figure 10: D-connector Insertion Stage
`
`correlation threshold
`tripped
`rotation to
`mate bayonet stubs
`
`depart move
`
`gripper opened
`
`28
`
`guarded move
`
`30
`
`32
`
`contact
`
`34
`
`rotation to lock bayonet
`
`40
`
`42
`
`44
`
`46
`
`48
`
`50
`
`52
`
`54
`
`36
`
`movement along -Y
`
`38
`
`time (sec)
`
`Figure 11: BNC coax Insertion Stage
`
`Page 7 of 8
`
`

`

`5 Conclusions
`
`7 References
`
`We have proposed a sensorimotor layer for easing the
`programming of sensor-based skills
`through sensor-
`integrated, task-relevant commands. The main idea is to
`encapsulate common uses of sensing and action for re-use
`on similar tasks within a domain. We have outlined several
`sensorimotor primitives in this paper and applied those
`primitives to sensor-based skills for performing connector
`insertions. Results from successful experimental trials were
`presented to show the feasibility of the ideas.
`Our initial goal of encapsulating sensor applications so
`they can be re-applied to different tasks was successful.
`However, the set of sensorimotor primitives is still very
`small, and a richer set must be developed and applied to a
`larger class of tasks. One problem with these connector tasks
`is the scale of contact is so small that it precludes detecting
`contact states with the force sensor (or by vision). These
`types of tasks usually require heuristic strategies, like those
`employed here, which are not guaranteed to succeed. In
`order to develop additional sensorimotor primitives, we
`need to identify larger-scale tasks (or improve our sensors)
`so that key task events can be adequately sensed. We intend
`to continue exploring the use of contact constraints for
`guiding force-driven primitive development. For vision
`primitives, we will explore the use of different feature types
`(e.g. edges) in order to implement some of the primitives
`discussed earlier.
`One of the significant research issues in this area is how
`to achieve generalization of sensor application. We are
`approaching it from a task-constraint point of view: both
`contact (force) constraints and visual constraints. To avoid
`making the primitives task-specific, constraints must be
`identified which are common across tasks within a domain.
`This, and other approaches to identifying task similarities,
`are the subjects of on-going research.
`
`6 Acknowledgements
`
`the
`in part by
`supported
`J.D. Morrow was
`Computational Science G raduate Fellowship Program of the
`Office of Scientific Computing in the Department of Energy.
`B. J. Nelson was supported in part by a National Defense
`Science and Engineering G raduate Fellowship through the
`U.S. A rmy Research Office through G rant Number
`DA A L03-91-0272 and by Sandia National Laboratories
`through Contract Number A C-3752D. We acknowledge
`Richard Voyles for helpful discussions. The views and
`conclusions contained in this document are those of the
`authors and should not be interpreted as representing the
`official policies, either expressed or implied, of the funding
`agencies.
`
`[1] R.W. Brockett, “On the computer control of movement,” Proc.
`of the IEEE Int. Conf. on Robotics and Automation, Philadel-
`phia, PA , 24-29 A pril, 1988, pp. 534-540.
`[2] J. Canny and K . G oldberg, “A RISC approach to robotics,”
`IEEE Robotics and Automation Magazine, vol. 1, no. 1, March
`1994, pp. 26-28.
`[3] D.C. Deno, R.M. Murray, K .S.J. Pister, S.S. Sastry, “Control
`Primitives for Robot Systems,” pp. 1866-1871, Proc. of IEEE
`Int. Conf. on Robotics and Automation, Cincinnati, OH, 1990.
`[4] B. Donald, “Planning multi-step error detection and recovery
`strategies,” Int. J. Rob. Res., vol. 9, no. 1, Feb. 1990, pp. 3-60.
`[5] M. Erdmann, “Understanding A ction and Sensing by Design-
`ing A ction-Based Sensors,” IJRR Draft, Nov 30, 1993.
`[6] A . Castano and S. Hutchinson “Visual Compliance: Task-
`Directed Visual Servo Control,” IEEE Trans. on Robotics and
`Automation, vol. 10, no. 3, June 1994, pp. 334-342.
`[7] S. Lee and H. A sada, “A ssembly of Parts with Irregular Sur-
`faces Using A ctive Force Sensing,” pp. 2639-2644, Proc. of the
`IEEE Int. Conf. on Robotics and Automation, San Diego, CA ,
`1994.
`[8] T. Lozano-Perez, “Task Planning,” Ch. 6, Robot Motion: Plan-
`ning and Control, MIT Press, Cambridge, MA , 1982.
`[9] T. Lozano-Perez, M.T. Mason, R.H. Taylor, “A utomatic Syn-
`thesis of Fine-Motion Strategies for Robots”, Int. J. of Rob.
`Res., Vol 3, No 1, pp. 3-23, Spring, 1984.
`[10]D. Lyons, “RS: A Formal Model of Distributed Computation
`For Sensory-Based Robot Control,” COINS Tech. Rept. 86-43,
`University of Massachusetts at A mherst, Sept. 1986.
`[11]G .H. Morris and L.S. Haynes, “Robotic A ssembly by Con-
`straints,” pp. 1507-1515, Proc. of the IEEE Int. Conf. on Robot-
`ics and Automation, Raleigh, NC, 1987.
`[12]J. D. Morrow and P.K . K hosla, “Sensorimotor Primitives for
`Robotic A ssembly Skills,” to appear in Proc. of the IEEE Int.
`Conf. on Robotics and Automation, Nagoya, Japan, May 1995.
`[13]B.J. Nelson, N.P. Papanikolopolous, and P.K . K hosla, Visual
`servoing for robotic assembly. Visual Servoing - Real-Time
`Control of Robot Manipulators Based on Visual Sensory Feed-
`back. ed. K . Hashimoto. River Edge, NJ: World Scientific Pub-
`lishing Co. Pte. Ltd. pp. 139-164.
`[14]W. Paetsch, and G . von Wichert, “Solving Insertion Tasks with
`a Multifingered G ripper by Fumbling,” pp. 173-179, Proc. of
`the IEEE Int. Conf. on Robotics and Automation, A tlanta, G A ,
`1993.
`[15]J. M. Schimmels and M.A . Peshkin, “A dmittance Matrix
`Design for Force-G uided A ssembly,” IEEE Trans. on Robotics
`and Automation, Vol. 8, No. 2, pp. 213-227, A pril, 1992.
`[16]T. Smithers and C. Malcolm, “Programming Robotic A ssem-
`bly in terms of Task A chieving Behavioural Modules,” DA I
`Research Paper No. 417, University of Edinburgh, Edinburgh,
`Scotland, December 1988.
`[17]D.B. Stewart, Schmitz, D.E. and K hosla, P.K . The Chimera II
`real-time operating system for advanced sensor-based control
`systems. IEEE Trans. Sys., Man Cyber. 22(6):1282-1295.,
`1992
`[18]D. Strip, “Technology for Robotics Mechanical A ssembly:
`Force-Directed Insertions,” AT&T Technical Journal, Vol. 67,
`Issue 2, pp. 23-34, March/A pril 1988.
`[19]W.O. Troxell, “A robotic assembly description language
`derived from task-achieving behaviors,” Proc. of Manufactur-
`ing International ‘90, A tlanta, G A , 25-28 March 1990.
`[20]D. Whitney, “A Survey of Manipulation and A ssembly: Devel-
`opment of the Field and Open Research Issues,” in Robotics
`Science, ed. M. Brady, MIT Press, 1989.
`
`Page 8 of 8
`
`