Page 51 - Human Inspired Dexterity in Robotic Manipulation

Sensorimotor Learning of Dexterous Manipulation 47

linkage causing a greater force-scaling bias than one presentation [44].
It should be emphasized that the lack of context cues and randomization
of object properties with the same visual appearance in these studies prevent
context-dependent error-based learning from occurring, therefore making
the use-dependent adaptation the dominant process influencing motor out-
put on the subsequent trials. For instance, it has recently been demonstrated
that the compensatory torque exerted by the subjects during trial sequences
with randomized target torque directions was biased toward the torque
direction experienced on the previous trial and with a magnitude of about
60 N mm [39]. By using the parameter obtained for our DPNI model, we
can simulate this process by setting the error-driven process to zero for all
trials, therefore obtaining a compensatory torque of 50 N mm, which is
in qualitative agreement with the experimental results of [39].
Our framework assumes the secondary sensorimotor process to be use-
dependent [16] rather than an error-based one with different time scales
[14]. This is because previous studies suggested that sensorimotor memory
could be established in the absence of a behavioral error [25]. Additionally,
such use-dependent adaptation of manipulation may be interpreted as a pro-
cess that encodes the likelihood of the occurrence of a context [40], similar
to the establishment of a fast-adaptive prior of the statistical properties of the
reaching movement independent of movement errors [17]. Use-dependent
plasticity has been recently proposed to be involved in the motor learning of
reaching movements [6,16]. Such use-dependent learning in movement
tasks is usually manifested as a bias in the movement direction. The encoding
of bias direction could be modeled as a set of Gaussian-like tuning functions
whose weightings are modulated by recent actions. Moreover, it has been
demonstrated that the directional bias had a larger influence on movements
made to angular targets further away from the repeated movement direction
[17]. This is consistent with our model, in which a greater bias was generated
when the subsequent target torque was in the opposite direction that had
been previously experienced. However, the use-dependent learning of
movement direction is often considered as a slow process that requires many
repetitions of the same action [15,16]. In contrast, our model seems to sug-
gest a rather fast process that requires only a few trials. We suspect that this
was due to fundamental differences between movement tasks and grasping
tasks like the one studied here. Specifically, movement tasks are usually
dynamic and require changes in muscle activation in a short time to accel-
erate and decelerate the hand/fingers. In contrast, object lifting tasks usually
feature a holding phase that requires subjects to produce a significantly large