Page 32 - Human Inspired Dexterity in Robotic Manipulation

28 Human Inspired Dexterity in Robotic Manipulation

Learning dexterous hand-object interactions, such as using a screwdriver
or modulating grip force on the handle of a hammer when hitting a nail over
consecutive times, relies on several sensorimotor processes. At the sensory
level, these processes involve multiple sensory modalities (i.e., vision, pro-
prioception, and touch), whose integration conveys information about the
state of the hand’s musculoskeletal system and the object in two ways. First,
sensory feedback is acquired and processed during hand-object interactions.
For review, see [2]. This informs the nervous system about object properties,
and the extent to which a given manipulation was performed as planned.
Second, sensory feedback is stored after hand-object interactions have been
executed, and used to anticipate object dynamics prior to performing sub-
sequent manipulations with the same or similar objects [3]. Thus, sensory
feedback about the current state of the hand and object, together with
sensorimotor memory from previous manipulations, are used to plan and
execute motor commands that are appropriate for the desired manipulation.
This general framework is consistent with the notion of internal models
through which the nervous system anticipates sensory consequences of a
planned movement to compute the corresponding motor commands [4].
Errors arising from a mismatch between desired and actual manipulation
outcomes would then be used to update such internal models in an iterative
fashion until the mismatch is eliminated.
A significant body of literature exists on human sensorimotor learning,
most of which has been focused on the control of arm movements using two
experimental approaches: adaptation of reaching movements against force
fields [5] and in response to visuomotor rotations [6]. Both approaches have
provided significant insights into humans’ ability to leverage a learned task
A to improve the learning rate of a similar task B, and conversely how
learning of task A might be—in certain circumstances—detrimental to
learning task B. For example, it has been shown that humans can generalize
the learned dynamics of a movement to neighboring directions to a greater
extent than with directions that are further away from the learned direc-
tion [7]. Contextual cues also appear to play a role in the extent to which
humans can generalize learned reaching movement dynamics [8].
In contrast to this work, much less is known about the mechanisms
underlying the learning of hand-object interactions, such as manipulation.
It should be emphasized that, due to the neural and biomechanical
differences between the arm and hand, one should not assume that
sensorimotor-learning theories developed using studies of reaching move-
ments would apply to the learning of manipulation tasks. Such differences