Page 38 - Human Inspired Dexterity in Robotic Manipulation

34 Human Inspired Dexterity in Robotic Manipulation

3.2.3.1 Data Analysis
As mentioned previously, we use T com to quantify the anticipatory control
of manipulation. In addition, we do not distinguish between contexts L and
R as we have already shown that there was no difference between the two
contexts from the perspective of torque production [31]. We have validated
the use of the T com to quantify the learning of high-level representations of
manipulations independent of the trial-by-trial variability of digit placement
and forces engaged in the task [32,36]. Note that the T com at object lift onset
is a direct measure of the anticipatory control of manipulation as it is com-
puted before the object is lifted. As such, the T com in Trial 1 in all protocols
was a measure of anticipatory control of manipulation based only on visual
geometric cues. The T com during the following trials was expected to be
influenced by both visual geometric cues and sensorimotor memory
acquired through manipulations performed on previous trials. Importantly,
the discrepancy between the T com produced at lift onset and the torque
required to perfectly prevent object roll (T target ) positively correlates with
the error in behavioral performance quantified as the object peak roll.
We define the sign of the required torque (i.e., T target ) based on the order
of the contexts (A or B), rather than the actual torque direction (CW or
CCW). Thus, contexts A and B were defined as being characterized by a
positive and negative T target , respectively. Moreover, the T com exerted by
the subjects would have the same sign as the T target if the direction of the
exerted torque was the same as the T target . For instance, if the T target in con-
text B was in the CW direction, the T com would be positive and negative if
subjects exerted a torque in the CCW and CW direction, respectively.
We were interested in quantifying the subjects’ ability to recall a learned
context or transfer to a new context (i.e., from context A to B; Fig. 3.1B).
To evaluate the retrieval of previously learned manipulation context, we cal-
culated the retrieval index (RI) as the difference between the T com exerted
on the retrieval trial and the averaged T com exerted across the last five trials
performed in the same context, normalized by the sign of the T target . For
instance, for the Ctrl group, the RI of context A was calculated as the dif-
ference between the T com exerted in Trial 1 of Block 3 and the T com aver-
aged across trials 4–8 of Block 1. Therefore, a negative or positive RI would
indicate an imperfect recall of a learned manipulation context, and the mag-
nitude of RI would correlate with the subjects’ inability to retrieve learned
manipulation. The assessment of transfer performance was based on the
within-subject comparison between the T com exerted in the transfer trial
and the T com exerted on the first trial in which subjects had no prior