Page 42 - Human Inspired Dexterity in Robotic Manipulation

38 Human Inspired Dexterity in Robotic Manipulation

3.2.3.3 Results
Learning manipulation in context A: first block of trials: All experimental groups
except the Rndm group started with the learning context A through a block
of eight consecutive trials. Based on our previous study [31], we expected
subjects to start with an estimation of the object weight distribution based
on visual geometric cues. Therefore, we expected subjects to generate
the compensatory torque (T com ) in the correct direction on the first trial
of Block 1. Through subsequent trial-by-trial learning, subjects would then
be able to quickly improve the estimation of the torque magnitude and its
timing within the first three trials. Our results are consistent with these
predictions using the same L-shaped object. Specifically, subjects started
by under-estimating the task torque (mean SE T com ¼ 182.25
9.41 N mm; n ¼ 84), but quickly approached the target torque within
the first three trials. Repeated measures ANOVA across the first three trials
revealed a significant main effect of Trial (P < .001) and no effect of Group
(P ¼ .732). Furthermore, repeated measures ANOVA across the last five
trials showed no effect of Trial (P ¼ .256) or Group (P ¼ .108). After the
learning context A in the first block, we examined how the preceding
manipulation context could affect the retrieval of a learned context or the
transfer to a new context by systematically varying the block sequence
and break time between certain blocks (Fig. 3.1B).
Control group: Interference occurs at both transfer trial and retrieval trial: Subjects
in the Ctrl group rotated the object 180 degree after every block of eight
trials, thus switching the manipulation context three times on the first day
(Fig. 3.1B). We have demonstrated that, when the manipulation context
(i.e., the direction of compensatory torque) was reversed after a block of
consecutive trials in context A, subjects made a significantly large error
on the first trial following the change to context B [31]. In the present exper-
iment, we found a similar result in the Ctrl group, despite the fact that the
change of contexts was induced by object rotation instead of changing the
handle of the U-shaped object that had to be grasped, as done in Fu and

Santello [31]. Specifically, on the transfer trial, subjects generated a T com
of 75.05 11.01 N mm which had a correct direction, but a significantly
smaller magnitude than the T com generated in Trial 1 of Block 1 (Fig. 3.2;
P ¼ .005). This suggests that subjects not only failed to generalize what they
had learned in context A to context B, but also performed worse than when
starting without prior manipulation experience, thus indicating a negative-
learning transfer.