Page 48 - Human Inspired Dexterity in Robotic Manipulation

44 Human Inspired Dexterity in Robotic Manipulation

process quickly generates a short-term “bias” through the performance of
manipulation within only a few trials (one to three) that competes with
the internal representation.
We first rule out models that suggest that memory of manipulation in the
first context is fragile and learning of the second context would overwrite
the previously learned manipulation in the first context (i.e., unlearned).
Such unlearning of the first context during learning of the second context
has been proposed to be due to either a single learning process [13],orto
be insensitive to contextual cues [5,34]. Specifically, these models predict
that a break given after the AB paradigm would not reduce the interference
on the recall of context A. We then considered the dual-rate multiple
context model [14], which supports the protection of a learned context
by assuming a context-independent fast process and a context-dependent
slow process. This model can successfully explain anterograde interference
as the fast process cannot be switched to the opposite direction on the first
trial following a context switch, therefore competing with the slow process
which is sensitive to contextual cues. However, we found that it is challeng-
ing for the DRMC model to account for our data from both the IF and RT
groups, although the DRMC model could reasonably predict the behavior
of Ctrl and Rndm groups. Specifically, the output of the DRMC model
results from the linear summation of the slow and fast processes, therefore
the decay of the fast component would cause a decrease of the interference
for the IF group, as well as a decrease of retention for the RT groups in a
similar fashion. This contradicts our data showing that the longer break
duration only significantly reduced the interference, but not the retention.
In fact, other models, such as the use-dependent learning model from [16],
that use linear combination of multiple sensorimotor processes when gen-
erating total motor output would have come to the same conclusion.
To better account for our data, here we propose a new model that
combines an error-based learning process and a use-dependent memory
(Eqs. 3.1–3.7) with six free parameters (A, B, C, D, E, F), which were deter-
mined using nonlinear optimization with data from the Rndm group, Ctrl
group, as well as the IF groups. By implementing the boot-strap technique,
we estimated parameters with 95% confidence intervals as follows:
A ¼ 0.9771 (0.9321, 0.9891), B ¼ 0.2613 (0.1848, 0.4435), C ¼ 0.4865
(0.2746, 0.7212), D ¼ 0.5513 (0.4619, 0.6410), E ¼ 0.0057 (0.0031,
0.0097), and F ¼ 13.37 (7.21, 28.19).
The new model also correctly predicted the data from the RT and TF
groups which were not used in the parameter search. Specifically, the new