Page 33 - Human Inspired Dexterity in Robotic Manipulation

Sensorimotor Learning of Dexterous Manipulation 29

include the hand’s greater number of degrees of freedom, and the fact that
manipulation provides richer contextual cues than reaching (i.e., object
geometry provides information about the object’s mass distribution), and
therefore the way it would react to forces and torque exerted by the hand.
This chapter reviews recent studies and our current understanding of
how humans learn to perform dexterous manipulation in the context of the-
ories of sensorimotor learning and concludes with open questions and future
directions for research on acquisition of dexterity in humans and robotic
applications.

3.2 LEARNING MANIPULATION: THEORETICAL
AND EXPERIMENTAL EVIDENCE
3.2.1 Background

Sensorimotor learning and memory have been studied extensively in motor
neuroscience using adaptation experiments in reaching and oculomotor
tasks [9]. In general, all skilled motor behavior relies on learning both control
and prediction which can be considered as two reciprocal procedures: control
generates motor commands to produce desired consequences, whereas pre-
diction maps motor commands (i.e., efferent copy) into expected sensory
consequences [10]. The mechanisms underlying these two procedures have
been traditionally proposed as inverse and forward internal models, respectively
[4]. The update of internal models is considered to be driven by the error
caused by a mismatch between sensed and predicted sensory outcome on
a trial-by-trial basis. This error is assumed to signal the direction of adjust-
ment for the next trial for error reduction. Error-based learning has been
demonstrated in many motor adaptation schemes, including saccade adap-
tation [11], visuomotor adaptation [12], and force-field adaptation [7].
State-space models are often used to describe error-based updates by defin-
ing the estimation of task dynamics as states that can be modified by errors
with learning rates. The exact structure of the model may vary as it could
consist of a single learning rate [13] or multiple rates [5], and could include
context selectors [14]. It has been demonstrated that these models can
capture many well-known phenomena including interference and savings.
Recent research has also demonstrated other sensorimotor learning pro-
cesses. Besides explaining additional performance improvements after errors
have been minimized, these nonerror based learning processes were also
found to be operating in parallel with error-based learning [6]. Reinforcement
learning uses signals, such as success and failure, which do not provide the