Page 46 - Human Inspired Dexterity in Robotic Manipulation

42 Human Inspired Dexterity in Robotic Manipulation

( 159.35 16.17 N mm) was significantly closer to the T target than the
T com in the transfer trial of the Ctrl condition (one-way ANOVA, P ¼ .001;
Fig. 3.3B). Additionally, a paired-sample t-test revealed that, for the TF60
group, the magnitude of the T com in the transfer trial was not significantly
different from the T com in the first trial of Block 1 (P ¼ .455). These findings
suggest that subjects started learning context B without any transfer from the
previously learned context A, even when a one-hour break was long enough
for the interference effect to decay to a minimum level. This phenomenon is
consistent with those reported by previous studies [27,29] which showed no
task-level learning transfer from context A to B. Furthermore, here we dem-
onstrate that a lack of learning generalization in manipulation tasks was not
due to lack of context cues or anterograde interference caused by learning a
context immediately preceding the new context.
Random group: One trial is sufficient to induce interference on manipulation in
the following context: The last experimental condition, Rndm, consisted of
pseudo-random presentations of the two manipulation contexts. This design
prevented subjects from being exposed to the same context for more than
three consecutive trials during the first half of the experiment. The first con-
text switch occurred after the first trial, instead of eight trials as in all other
groups (Fig. 3.1B). We found that subjects in the Rndm group exerted a
torque (T com ¼ 138.33 22.52 N mm) closer to the T target than the tor-
que exerted by the Ctrl group on the first trial of context B. One-way
ANOVA confirmed a significant main effect of the group (P ¼ .011). Addi-
tionally, the magnitude of the T com subjects exerted for the first trial of con-
text B was significantly smaller than the T com exerted for Trial 1 of context
A (repeated-measure ANOVA, P ¼ .028). These two findings suggest that
one trial was sufficient to induce interference in the next context, whereas
further practice in the same context would have increased the strength of the
interference.
The Rndm condition also required subjects to frequently switch the
manipulation context multiple times during the first half of the experiment.
We found that the performance on the context switch trials was generally
worse than preswitch trials, despite the fact that the manipulation in the pre-
switch context was performed for no more than three consecutive times. To
illustrate this, for each context, we compared the average of the last three
preswitch trials (Trials 5, 12, and 16 for Context A; Trials 9, 10, and
14 for Context B), the average of the last three context switch trials
(Trials 7, 11, and 15 for context A; Trials 8, 13, and 17 for context B), as
well as the average of the last three blocked trials (Trials 30–32 for context