Page 53 - Human Inspired Dexterity in Robotic Manipulation

Sensorimotor Learning of Dexterous Manipulation 49

Parallel representations of dexterous manipulation: We have described how
learning dexterous manipulation in one context may be penalized by learning
another context, even when the target object was the same across both con-
texts. This was due to the interaction between context-independent and
context-dependent learning processes. Context-dependent representation
is advantageous when the environment/object provide clear context cues
which assist recall of context-dependent memory. In contrast, context-
independent representation was developed based on the “history” of events,
and it could be useful when the upcoming context is unpredictable due to the
lack of context cues. We think that it could be also beneficial to implement
multiple learning mechanisms working in parallel during robotic learning of
manipulation tasks, as this may be used to deal with uncertainty in the envi-
ronment. Furthermore, in robotics, one has the opportunity to design the
controller from scratch. Therefore, the negative effect between multiple
learning processes described for human sensorimotor learning could be min-
imized when such parallel structure is implemented artificially in robots.
Digit force-to-position modulation: In this chapter, we did not focus on
how each individual finger is controlled during fine manipulation tasks.
However, we have reported strong evidence in other publications that
humans have the ability to modulate manipulative forces as a function of
where the object is grasped. We have reported that such modulation, occur-
ring within a few hundred milliseconds from contact, is critically important
for ensuring that the manipulation goal is attained despite a trial-by-trial
variability in digit placement [36]. We also showed that such digit force-
to-position modulation is found when subjects are asked to perform a given
manipulation task on the same object grasped with a different number of
digits [32]. These findings underscore an important phenomenon: Subjects
may not learn a rigid mapping between digit positions and forces, but rather
build a “high-level” representation of the task (e.g., a given compensatory
torque) that the nervous system is then able to use to coordinate effectors in a
flexible manner depending on environmental constraints, (e.g., number of
digits available to perform the task or unpredictable variability of digit place-
ment). This phenomenon also implies that the system is able to (1) sense digit
placement online, (2) communicate this feedback to the high-level task
representation, and (3) modulate forces accordingly. One could therefore
envision a robotic analogue, whereby learning—by trial and error— the task
that high-level representation could potentially be used to drive dexterous
manipulation using different hardware, (i.e., manipulators with different
number of joints or digits).