Page 34 - Human Inspired Dexterity in Robotic Manipulation

30 Human Inspired Dexterity in Robotic Manipulation

gradient of error reduction, therefore being exploratory to some extent.
Rewards or penalties are given to a certain action that might have led to
the success or failure and eventually lead to optimal solutions [9]. Another
learning process has been termed use-dependent learning that describes the
ability to adapt motor commands through repetition of movements without
outcome information [15]. There is evidence that use-dependent learning
could influence the directional control of subsequent reaching movement
[6,16,17].
Sensorimotor learning and memory can also be evaluated from the per-
spective of retention and generalization. Generalization tests the extent to
which a learned task could benefit other tasks that may share characteristics
with the learned task. It has been shown that sensorimotor learning might be
context dependent, and that the ability to generalize to other tasks follows a
bell-shaped tuning curve depending on the extent to which the new context
is similar to the learned one [18]. Retention requires subjects to recall a
learned task after a break of various durations. A learned motor task can
be retained for at least 24 h [19,20]. It has also been shown that learning
a secondary task in the opposite direction interferes with the retention of
the first learned task [21], although the mechanisms underlying protection
and retrieval of motor memory remain unclear [22].
Despite the success of the previously mentioned computational frame-
works in capturing experimental observations, sensorimotor learning of
multidigit dexterous manipulation tasks has received little attention.
Specifically, it remains unclear how the central nervous system (CNS) learns,
stores, and retrieves knowledge about previously performed manipulations.
Most research in the grasping literature has been focused on empirical find-
ings about force modulation to an object’s weight to enable the lifting of an
object. When subjects initially misjudged the weight, trial-by-trial learning
was usually assessed by the improvement in force generation (peak force rate
modulation to object weight) that most closely matched the actual weight.
There is evidence that the internal representation of force scaling could
be updated, maintained, and retrieved given appropriate visual cues
[10,20,23,24]. In addition to memory associated with a specific task, it
has also been demonstrated that task-independent memory could be formed
and influence the subsequent manipulation [25]. A novel experimental
protocol using objects with asymmetrical mass distribution [26] has been
developed to provide more insights about how manipulation is learned
and represented. In this protocol, task performance can be assessed by errors
made during lift because subjects are required to lift the object straight and