IEEE Robotics & Automation Magazine - December 2022 - 76

(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure 8. Human demonstrations and robot force and impedance adaptation control. (a)-(c) A demonstration through kinesthetic
teaching. (d) The teaching result. (e)-(g) Robot force and impedance adaptation to track the demonstrated trajectory. (h) The
trajectory adaptation results.
Figure 8(a)-(d) shows the process and final drawings of
demonstration. The human operator handled the pen to draw
a square and the robot arm recorded the position while drawing.
Then the robot tried to follow the human demonstration
by increasing the feedforward force and impedance factors.
Here, we use the robot to record the trajectory in one trial and
send the data to the simulation environment to generate the
force and impedance for the next time to complete the trajectory
tracking in the unstable force field. In Figure 8(e) and (f),
the robot tried several times to finally follow the trajectory
demonstrated by human operator.
Discussion
The proposed learning and control framework transfers iterative
learning control into a fully NN, and the neural node
linkage is endowed with biomimetic meanings, i.e., the
impedances are affected by the feedforward torques by adding
several enhancement nodes that have nonlinear relationships
with the feature mapping functions. The aforementioned
quantized simulation results show that the proposed framework
has a faster convergence rate to the desired effect, a
more robust response to the different environmental conditions,
and better continuity and uniformity of the feedforward
force in a complex force field. Nevertheless, there are some
improvements for the current framework. For example, the
76 * IEEE ROBOTICS & AUTOMATION MAGAZINE * DECEMBER 2022
learning rates of torques and impedances are not controllable,
while in [8] they can be changed by choosing appropriate gain
matrices. On the other hand, the difference between the force
drop in the simulation and the force drop observed in reality
is difficult to diminish through this framework. Today, deep
learning methods can transfer motions and parameters from
human demonstrations to robot manipulation [15], [16] to
minimize the error between simulation and reality. In the
future, it is desirable to combine the research of this work
with deep learning methods to develop a system for human
skill learning and generalization.
Conclusion
In this article, a new biomimetic learning and control framework
for force and impedance was proposed based on the
theory of BLS. Following the CNS and human motor working
mechanism, an incremental and explainable NN with
functional linkage was constructed for force and impedance
adaptation and trajectory tracking in stable and unstable
force fields. Some principles, such as the impedance of muscle
fibers and amount of activated muscle fibers were reflected
by enhancement nodes in a flat NN, which have hardly
been modeled in previous research. Three groups of comparative
simulations based on the open access data set
were performed to verify the merits of tracking speed and

IEEE Robotics & Automation Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2022