Page 52 - Human Inspired Dexterity in Robotic Manipulation

48 Human Inspired Dexterity in Robotic Manipulation

constant torque/force in addition to a short object lifting movement. This
isometric force production may play an important role for building up the
motor bias. Furthermore, reaching could be potentially executed toward an
infinite number of continuously distributed directions, whereas object-
balancing tasks are bi-directional tasks (CW and CCW) whose performance
relies on the distribution of force magnitude across agonist and antagonist
muscle groups. The binary choices of torque direction might make our task
more deterministic, therefore leading to faster adaptation.
Our study examined retention and interference of manipulation using
tasks that require torque production and suggest that interference occurring
at context switch was due to conflict between two sensorimotor processes.
However, as the use-dependent component decays fast, retrieval of learned
manipulation was successful even after 2 weeks. Interestingly, the interfer-
ence found when switching context did not completely disappear after mul-
tiple switches. This indicates that even strong visual geometric cues could
not completely override the effect of preceding manipulations. However,
visual cues about object weight have been shown to completely override
the effects of previous lifts on grip force scaling [45,46]. This can be
explained by the subtle difference between grip force scaling and torque
production. Specifically, while torque generation in object balancing is
bi-directional, the grip force scaling in object lifting is essentially unidirec-
tional. Our nonlinear bias framework indicates that the bias from the pre-
vious motor action could be quite small if the subsequent action is
performed in the same direction. Therefore, bias from previous-object lift-
ing may not be strong enough to cause a measurable difference when visual
cues about an object’s weight is provided. Despite this difference, our result
is consistent with early grasping studies showing that memory of object
weight distribution with context cues could last >24 h [20,24].

3.3 LESSONS LEARNED FROM HUMAN DATA AND
POTENTIAL APPLICATIONS TO ROBOTIC DEXTEROUS
MANIPULATION

The previously reviewed human data, as well as other data from our labo-
ratory not only provide important insights on sensorimotor learning and
control of human dexterous manipulation but could also open new research
avenues for robotic dexterous manipulation. We hereby summarize two of
these robotics research avenues.