Page 50 - Human Inspired Dexterity in Robotic Manipulation

46 Human Inspired Dexterity in Robotic Manipulation

slower time scale [16]. Finally, the parameter F suggests that, when not
actively involved in the task (i.e., during break), the use-dependent
memory u decays with a half-life of 9.27 min with a 95% confidence interval
(4.99, 19.54).

3.2.3.4 Discussion
There has been extensive evidence that the current manipulation could be
influenced by previous manipulations. Specifically, it has been shown that
the CNS is capable of a fast trial-by-trial adaptation to changes in manipu-
lation context, for example, object weight [41], surface friction [42], and
weight distribution [36]. Such adaptations are thought to rely on a
“sensorimotor memory” which was initially defined as the memory of an
object’s physical properties [3]. However, more recent data suggested that
the CNS may maintain multiple internal representations of manipulator
skills. Quaney et al. [25] demonstrated that squeezing an object generates
a bias to grip force for object lifting in the following trial, and that the
grip-force bias induced by a strong pinch disappeared when lifting the object
after 24 h [43]. These results suggested that there might be an object-based
memory for object weight, and an action-based memory for pinch force.
Our findings provide direct evidence that support the concept that multiple
sensorimotor mechanisms underlie the fast trial-by-trial adaptation
(i.e., establishment of the “sensorimotor memory”) of object manipulation.
Most importantly, we revealed that one component of these adaptation pro-
cesses lasts a relatively short time while the other could be correctly recalled
after a long period of time. Furthermore, these two mechanisms with
different time scales interact during the generation of motor commands
for subsequent manipulation, thus inducing interference when they were
incongruent with each other.
Our experimental design used an object that provides strong visual geo-
metric cues about the properties of the object, which promotes context-
dependent learning as well as visually based motor planning. Interestingly,
we also found that our theoretical framework can be used to explain unpre-
dictable trial conditions when context cues are absent. In simple object-
lifting tasks, digit forces are scaled to the object weight experienced in
the previous trial if information about object weight is unavailable on the
current trial [38]. In bimanual manipulation tasks where subjects cannot reli-
ably predict whether the object is mechanically linked between two hands,
the scaling of the finger forces depends on the presence of the linkage expe-
rienced in the previous trials, with three consecutive presentations of the