Robotic object grasping in context of human grasping and

manipulation
Pavel Dzitac, and Abdul Md Mazid, Member, IEEE

Abstract— This paper presents experimental and grasp force safety margin as small as 10% while still
deductive findings that shed new light on grasp force manipulating objects safely, is based on a predictive
estimation, which improves robot’s chances to grasp control model of the manipulator (hand and arm) and the
and manipulate the object close to optimum conditions object. In this grasp force estimation model the role of the
on the first attempt, which in turn improves robot’s sensory system is to provide feedback to the predictive
object manipulation dexterity. model [2], necessary for immediate corrections and to
This paper proposes that object slippage detection in improve estimation performance by learning.
the human hand is not detected based purely on micro- The required sensory feedback in this grasp force
vibrations sensed by the human skin during incipient estimation model is the normal grasp force sensing,
slippage but also on load sensing at each finger and tangential force sensing between the object and the gripper
movement of fingers relative to each other while and slippage detection.
holding an object. The normal grasp force and tangential force sensing are
used for dynamic grasp force control necessary to apply
and maintain the predicted grasp forces. The slippage
Keywords – slippage, robotic grasping, grasp force, human detection function is required to trigger a correction to the
grasping. predictive model whenever the predicted grasp forces are
insufficient to prevent slippage, such as during unexpected
environmental conditions [3].
I. INTRODUCTION
Researchers revealed that people learn to predict the
I NDUSTRIAL robots manipulate objects well because
they are equipped with dedicated technology and
manipulators designed for known objects whose mass,
consequences of their actions before they learn to control
their actions [4]. A person can predict how a ball should be
shape, sizes, material, surface texture and other relevant struck well before learning how to do it like a professional.
object characteristics are known in advance. Generic This also implies that to be able to predict or control its
object manipulation adds an additional degree of actions, the manipulator must be capable of learning.
complexity to object manipulation due to the need to Manipulators must also be able to recognise and compare
estimate object characteristics on the fly, as in the case of objects in order to make prediction based on similarity
unknown object grasping [1]. possible. The overall predictive model is therefore based
on the ability to learn, ability to make predictions based on
Neurological research suggests that humans are equipped learned knowledge and ability to extract information from
with two internal control models: a forward (predictive) sensory feedback, which may be used to confirm
model and an inverse (reactive) model [2]. Both of these predictions or trigger corrective actions when predictions
control models result in movement of human body. The prove to be incorrect [3]. In other words, the grasp force
predictive model predicts future required actions, while the prediction model is based on sensory feedback and
reactive model responds to sensory feedback. When intelligence.
combined, the two control models are similar to the well
known closed-loop control model that adjusts the control For the purpose of industrial robotic object manipulation,
output based on feedback from sensors, but can also where the object characteristics are known in advance, the
contain a “feed-forward” loop to adjust the control output predictive model could be used to enhance manipulation
in anticipation of future requirements. dexterity by reusing grasp force information to manipulate
similar objects. However this requires the robot to see and
Neurological research also suggests that the human grasp sense the object and be equipped with an autonomous
force estimation ability, which manages to maintain a model that can predict object characteristics based on
similar known objects.
Manuscript received June 10, 2013.
This paper looks at object parameters and their influence
Pavel Dzitac is with the School of Engineering and Technology,
Central Queensland University Australia, Rockhampton, Australia on grasping, the importance of grasp force safety margin,
(phone: +613 92117615; email: pavel.d [email protected] m). the mechanics of object slippage and the sources of
Abdul Md Mazid is with the School of Eng ineering and Technology, information during object slippage.
Central Queensland University Australia, Rockhampton, Australia
(phone: +617 4930 9672; email: [email protected]).

978-1-4799-1201-8/13/$31.00 2013
c IEEE 201
 II. GRASP FORCE AND SLIPPA GE CONT ROL Centre of gravity:
A. Object Parameters and Characteristics x Remains constant for rigid objects during
manipulation;
From the perspective of object grasping and manipulation
x Changes for deformable objects during
the “customer” is the object itself. Therefore it is worth
manipulation;
taking into account the influence that various object
x Influences the required tactile sensor load
parameters and characteristics have on object grasping and
capacity;
manipulation.
x Influences the required tactile sensor complex
Object density, volume, shape, material type and surface load sensing capacity (tangential and torsional);
texture are important object parameters that can be varied x Influences the required grasp force capacity;
to obtain specific object characteristics such as mass, size, x Influences manipulation speed;
centre of gravity, coefficient of friction, rigidity and x Limits range of objects that can be grasped and
fragility [1]. Each of these characteristics influences object manipulated with a particular gripper;
grasping and manipulation in its own way. x Limits range of objects that can be grasped and
The following are some of the important behaviours and manipulated with a particular manipulator;
influences that object shape, mass, size, centre of gravity, x Can be used to an advantage to reduce grasp
coefficient of friction, rigidity and fragility have on object forces and increase grasp reliability;
grasping and manipulation.
Coefficient of friction:
Shape: x Influenced by the coefficient of friction of the
x Remains unchanged for rigid objects during gripper contact surface;
manipulation; x Influenced by surface texture;
x Changes for deformable objects during x Influenced by temperature and humidity;
manipulation; x Influenced by object surface contaminants;
x Influences grasping and manipulation x Influences slippage;
complexity; x Influences required grasping forces;
x Limits range of objects that can be grasped and
manipulated with a particular gripper; Rigidity:
x Can be used to an advantage to reduce grasp x Influences grasping complexity;
forces and increase grasp reliability; x Influences the sensitivity requirements of tactile
sensors;
Mass: x Can be used to an advantage to reduce grasp
x Remains constant during manipulation; forces and increase grasp reliability;
x Influences robot lifting and manipulating
capacity; Fragility:
x Influences manipulation speed and inertia; x Influences grasping complexity;
x Influences robot grasp force capacity; x Influences the sensitivity requirements of tactile
x Influences the required tactile sensor load sensors;
capacity; x Influences the mechanical requirements of tactile
x Limits range of objects that can be grasped and sensors;
manipulated with a particular gripper; x Influences precision grasping requirements;
x Limits range of objects that can be grasped and
manipulated with a particular manipulator; The complex behaviour and influences that object shape,
mass, size, centre of gravity, coefficient of friction and
Size: rigidity have on object grasping and manipulation suggest
x Remains unchanged for rigid objects during that object grasping and manipulation is a complex task.
manipulation;
x Changes for deformable objects during B. Grasp force safety margin
manipulation;
Given the potentially very large variability of object
x Influences the required gripper capacity;
characteristics from one object to another it can be
x Limits range of objects that can be grasped and concluded that even a highly sophisticated grasp force
manipulated with a particular gripper; predictive system, such as the human one, cannot be
x Limits range of objects that can be grasped and expected to predict with 100% success rate in all possible
manipulated with a particular manipulator; circumstances. Therefore the predictive system needs a

202 2013 6th IEEE Conference on Robotics, Automation and Mechatronics (RAM)
trigger mechanism that would request corrective actions Tangential forces at the gripper-object interface drop when
whenever the predicted grasp forces are dangerously low. slippage starts. This slippage detection strategy was less
This is where slippage detection becomes extremely reliable than feature movement across the sensor but gave
important. acceptable results.
One important question still remains though: what about
the excessive grasp forces? Research suggests that the The slippage detection strategy based on vibration (rapid
predictive grasp control model “probes” the validity of the tangential force changes) generated during bulk slippage
required grasp force “safety margin” via slippage detection was less reliable than force changes when slippage starts
feedback and the model adjusts it to some sensible value but still gave usable results.
[3].
The most difficult to detect and therefore the least reliable
No modality is suggested for correction of excessive grasp (but potentially very useful) was found to be incipient
forces. The human object manipulation model has a well slippage detection based on force changes at some sensor
known grasp force safety margin adjustment modality: the elements caused by changes in tangential forces during
feedback received when the manipulated object is incipient slippage. This was partially due to the limitations
damaged as a result of excessive grasp forces during of the tangential force sensor used (based on strain
manipulation. However, it is expected that it would be gauges). However, the difficulty of obtaining clear
difficult to implement such autonomous capability in a vibration signals indicates that in order to be useful, an
robot. incipient slippage detection sensor has to be highly
specialised, and potentially contain intelligent electronics
capable of extracting reliable incipient slippage
C. Slippage Detection and Prevention
information from a vibration-noisy source;
Experimentation was conducted using tactile sensor shown No attempt has been made during experimentation to
in Figure 1 to identify slippage events that could be monitor the change in contact area during incipient and
reliably detected and therefore used to detect slippage. bulk slippage, because the sensor used (Figure 1) did not
have sufficient spatial resolution.

D. Stick-slip Variability
Unlike object mass that does not change during object
manipulation, it has been observed that vibration
characteristics (stick-slip effect) changed with the object
Figure 1. Slippage detection sensor
being manipulated, the object surface conditions and
Flat steel plate of 200g, 400g and 600g was placed on the slippage speed [5]. The existence of a detectable stick-slip
sensor and then pulled until slippage occurred while effect depended on the right condition being present at the
monitoring force changes at each sensor element (bump). right time. Authors’ opinion is that although the stick-slip
It was found that the following events are detected when effect is a useful approach for slippage detection under
slippage occurs: right conditions, it may be an unreliable source of
x Force changes at some sensor elements occur information for incipient slippage detection under all
during incipient slippage due to changes in the conditions and for all materials.
tangential forces at the sensor-object interface;
x Tangential forces at the sensor-object interface E. Sources of Information during Object Slippage
drop when bulk slippage starts; A human-based experiment was conducted in an attempt to
x Vibration occurs during bulk slippage due to understand the sources of information that are used to
rapid changes in tangential forces (stick-slip) at detect and prevent slippage when humans manipulate
the sensor-object interface; objects.
x Object features (e.g. a hole in the object surface)
moves across the sensor array;

The most reliable slippage detection strategy was obtained

by monitoring the movement of an object feature across
the tactile sensor array. However such a feature may not
exist on an object substantially enough to be detected,
which makes this slippage detection strategy impractical
for generic object manipulation;
Figure 2. Slippage prevention experiment – split block and force sensor

2013 6th IEEE Conference on Robotics, Automation and Mechatronics (RAM) 203
 For the purpose of this experiment a 300 gram split-block, This suggests that when humans use multiple fingers to
fitted with a force sensor between the two split-block grasp an object they do not rely solely on incipient
halves was used (Figure 2). slippage detection at the finger-level, but also on
information like finger position and finger load change
Randomly selected subjects aged between 15 and 26 were relative to other fingers. This agrees with the view that
challenged to achieve the lowest possible grasp force tactile sensing should not be considered in isolation [6].
without allowing fingers to slip on the object. Unknown to
the subjects, the real purpose of the experiment was to Each of the human hand fingers in contact with the object
determine whether the incipient slippage event would provides resistance to slippage and contributes to slippage
allow the subjects to prevent slippage from occurring at detection and prevention.
all. The following were concluded from this experiment: High speed video demonstrates that when slippage starts
in a multi-finger grasp, not all fingers loose contact with
1. All subjects were able to reduce the grasp force from the object at the same time: some fingers begin to slip
a safe grasp force margin to the point where slippage while others still maintain grasp, as displayed in the video
occurred, which suggested that subjects were initially sequence in Figure 3.
holding the object with a grasp force safety margin of
about 20 to 30%; It is noticeable in Figure 3 (frame 2) how the index finger
2. None of the subjects could reduce the grasp force to a slips from its initial position and then instinctively grasps
point where they were certain that slippage is about to the object at a new location. The other fingers remain
begin, and therefore would stop grasp force reduction relatively unmoved (apart from a slight roll of fingers on
before finger slippage occurred. Incipient slippage the object).
may have been detected, but human conscious This grasp control mechanism behaves like the finger level
reactions may be too slow to prevent slippage based incipient slippage, but actually takes place at the hand and
on the incipient slippage event; arm level, and allows the human hand to prevent object
This agrees with the predictive grasp control model [3], bulk slippage by adjusting the grasp force of fingers before
which proposes that the human grasp force control model the object can start to slide downward.
maintains a grasp force “safety factor” above the required
grasp force (learned from experience), and only uses The slippage sequence in Figure 3 shows a dynamic
slippage events to readjust the control model. behaviour that is clearly visible in slow motion and has the
following components:
During further experimentation it was noticed that 1. Discrete finger slippage event;
multiple grasp contact points create a stick-slip macro 2. Slippage prevention action;
model at hand level similar to the localised stick-slip 3. Load-holding force adjustment action;
micro-model at finger-level.
The slippage event in Figure 3 is likely detected as
follows:
1. Initially the arm and hand are in a stationary position.
The grasp force is being reduced but no slippage takes
place yet;
2. Index finger slippage occurs (image 2);
3. Index finger moves relative to other fingers during
slippage, which together with the tangential load
reduction at the index finger and increase at the other
fingers is registered as slippage and likely takes place
in parallel with vibration at the fingertip due to stick-
slip;

Based on this sequence of events it could be argued that if

slippage prevention was based on vibration caused by
incipient slippage alone, the object should have been held
with some minimum grasp force without allowing bulk
slippage to occur at any of the fingers. However, the
evidence shows otherwise.
This result agrees with the concept of a predictive human
grasp force control model and the need for a grasp force
Figure 3. Slippage Control by Hu man Hand (600gram Ob ject). The safety margin. After all, if humans could reliably detect
images are 24ms apart

204 2013 6th IEEE Conference on Robotics, Automation and Mechatronics (RAM)
and prevent slippage based on vibration generated during measured at the finger that slipped was about 30% larger
incipient slippage there would be no broken dishes in the before slippage than the tangential force on the finger that
kitchen. did not slip. When the finger slipped, the measured
tangential force increased at the other finger where a lower
The slippage prevention action in Figure 3 takes place as grasp force was recorded before slippage. The result was
follows: unexpected because the experiment was suggesting that
1. Slippage is detected; the tangential forces developed at the two fingers were not
2. Contraction of all fingers is noticeable, meaning that related to the applied grasp forces at the two fingers.
all fingers increase their grasp force.

The load holding force adjustment takes place as follows

(notice the changes in hand position relative to the
background feature shown by the arrow in frame 1):
Force
1. The forearm rotates slightly upwards at the elbow at sensor
first due to the difference in the sensed weight and the location
original holding force;
2. Finger slippage causes the object weight share of the PTFE
inserts
slipping finger to be distributed to the other fingers;
3. This causes the fingers to “sag” from their original Figure 4. Slippage experiments with normal and tangential fo rce
position relative to the palm; sensing
4. Finger “sag” causes the hand to rotate downward at
It was found that the random finger slippage in a multi-
the wrist, which in turn causes forearm downward
finger grip can be explained as follows. For simplicity
rotation at the elbow (elbow is rested on a stool);
assume that a 1Kg object is being held in a three-finger
grasp, as shown in Figure 4. If the coefficient of friction μ
Experiments revealed that this behaviour is more visible
between the object and fingers is 0.5, a normal force of
with heavier objects and objects with higher coefficient of
minimum 10N on each side of the object is needed to hold
friction. Very light and slippery objects do not provide
the object. The tangential friction forces Ff1 and Ff2 that
good tangential force and stick-slip feedback, and
are required to hold the object can be calculated as
therefore readily slip out of control when grasp forces are
follows.
too low.
Ff1 = μF1 = 0.5*5N = 2.5N (1)
From the dynamics point of view, the initial upward
movement of the hand during slippage causes undesirable
Ff2 = μF2 = 0.5*5N = 2.5N (2)
acceleration in the opposite direction to object slippage,
which actually makes slippage more difficult to bring
Equations 1 and 2 suggest that the two tangential forces
under control. There isn’t much one can do about this
behaviour in the human object manipulation model. Ff1 and Ff2 must be equal because the grasp forces and the
However this undesirable behaviour could be minimised in coefficients of friction at the two fingers are equal. This in
the robot control model by moving the arm in the direction turn suggests that both fingers share the object holding
of slippage, if possible, in order to reduce the relative load equally. However, it turns out that this is only true if
velocity between the object and the gripper, and therefore both fingers offer the same stiffness (moment) to the
tangential (vertical) load.
reduce the time and force needed to stop the slippage.

As discusses earlier, not all fingers slip exactly at the same

time. Intuitively it was assumed that this was due to slight
different coefficient of friction at each finger.
A set of grasp force experiments was conducted using the
device in Figure 4. The experiments were conducted with
the index and the middle finger being located on PTFE
inserts that were fitted with sensors capable of measuring
the approximate normal and tangential grasp forces
simultaneously.
The grasp force was reduced until slippage on one finger
occurred. No significant difference in the grasp force
between the two fingers was noted; the difference was less
than 5%. However, it was noticed that the tangential force Figure 5. Finger Slippage Model

2013 6th IEEE Conference on Robotics, Automation and Mechatronics (RAM) 205
 This is because the friction forces Ff1 and Ff2 at the finger- be based purely on incipient slippage because human
object interface will only be effective in holding the object reflexes are likely to be too slow to react in time to prevent
if fingers develop a moment that prevents finger rotation, slippage.
as shown in Figure 6. This can be prevented in robotic object manipulation
because researchers have shown that a robot with fast
reactions can prevent bulk slippage based on incipient
slippage events [7]. However, a grasp force safety margin
increases grasp reliability and is therefore a more realistic
approach to safe object manipulation.

Experimental results also suggest that the concept of a

predictive grasp force control that maintains a grasp force
safety margin and uses slippage events as a correction
trigger is a reasonable fit to the experimental results in this
paper.

It has been established in this research that the main

sensors that the predictive grasp force model requires are
1. Normal and tangential force sensors, required for
Figure 6. Finger mo ments M 1 and M 2 dynamic grasp force control;
2. Slippage sensors required as a trigger for corrective
If finger 2 had a moment M2 = 0 it would not contribute to
action to the predictive grasp force model to allow
holding the object. This means that finger 1 has the whole
dynamic corrections whenever the predicted grasp
share of the object load. However at the original finger 1
forces are too low and to facilitate predictive model
grasp force of 5N, finger 1 would slip when the tangential
improvement by learning from experience.
force exceeds 2.5N unless it increases its grasp force to
achieve the required tangential friction force to hold the
In industrial robotics, when sufficient object
object, or finger 2 regains its stiffness and contributes to characteristics such as the mass, shape, size, coefficient of
the lift. friction and centre of gravity of an object are known, the
grasp and manipulation forces could be readily calculated
It was concluded from these simple experiments that in a and applied dynamically without the need to sense
multi-finger grasp, the fingers that hold a higher share of slippage. However this may not result in optimum
the load slip first. It was also concluded that when holding application of grasp forces under unpredictable conditions,
an object, equal grasp force at all fingers does not and therefore slippage detection can be used to trigger a
necessarily translate into equal load sharing. correction to the predicted grasp force model.
This information could be used to improve robot grasp
reliability when performing a friction-based grasp by
optimising the load sharing between the robot fingers and
therefore minimising chances of finger slippage, which REFERENCES
could create a slippage chain-reaction. [1] M. A. Mazid and R.A. Russell, “A Robotic Opto-tactile Sensor for
Assessing Object Surface Texture”, in Proc. IEEE Int’l Conf. on
III. CONCLUSION Robotics, Automation and Mechatronics (RAM 2006). Bangko k,
pp. 387-391, 2006.
The paper is a source of philosophical and technical [2] A.G. W itney, A. Wing, J.L. Thonnard and A.M. Smith, “The
analysis that facilitates understanding of human object cutaneous contribution to adaptive precision grip”, Trends
Neurosci. 27, 637, 2004.
grasping and how it relates to robotic object grasping. [3] A.M. Wing, J.R. Flanagan, and J. Richardson, “Anticipatory
postural adjustments in stance and grip”, Exp. Brain Res. 116, 122–
In this paper research by others, which suggests the 130, 2007.
[4] J.R. Flanagan, P. Vetter, R.S. Johansson, and D.M. Wolpert,
existence of a grasp force predictive model in humans, has “Prediction precedes control in motor learning”. Current Biology,
been presented. This is an important concept that explains 13(2):146–150, 2003.
how humans predict and control grasp forces, which [5] V. L. Popov, “Contact Mechanics and Friction: Physical Principles
contributes to the impressive human ability to manipulate and Applications,” 1st ed. Springer-Verlag : Berlin, Heidelberg,
2010.
objects skilfully and safely. [6] R. S. Dah iya, G. Metta, M. Valle and G. Sandini, “Tactile Sensing -
Fro m Hu mans to Humanoids”, IEEE Transactions On Robotics,
Experimental results lead to the conclusion that finger Vo l. 26, No. 1, February 2010.
[7] M.E. Tremblay and M.R. Cutkosky, “Estimat ing friction using
slippage in a multi-finger grasp may take place even when incipient slip sensing during a manipulation task”, In Proc. ICRA
the intent is to consciously prevent slippage from 93, volu me 1, pages 429– 434, 1993.
occurring. This suggests that slippage prevention cannot

206 2013 6th IEEE Conference on Robotics, Automation and Mechatronics (RAM)