IEEE Robotics & Automation Magazine - December 2022 - 67

fields to verify advantages of the proposed framework in
terms of force and trajectory tracking efficiency and
accuracy, robust responses to different force situations, and
continuity of force application in a mixed stable and unstable
environment. Finally, an experiment is conducted to verify
effectiveness of the proposed method.
Introduction
Variable force and impedance control mechanisms have
widely existed in human daily movements and work, such as
drilling and grasping, requiring the central nervous system
(CNS) to control muscle reflexes and learning processes
against intrinsic instability caused by unknown external forces,
motor noises, and raw material properties when interacting
with objects in our environment. In recent years, research
on motor skills learning has been used for robotic applications,
such as teleoperation [1], transferring skills from
humans to robots [2], and haptic interaction [3], [4], integrating
electromyography perception and learning technology.
To realize the human motor learning process, a mechanism
used by the CNS is proposed, involving both feedback
and feedback control, to represent the relationship between
motor commands and movement so that the CNS can adapt
the dynamics of the limb to compensate for mechanical instability
and interaction forces from the environment [6]. However,
even based on the common mechanism, there are some
models that describe human motor control and learning, but
they are slightly different from each other. Franklin et al. built
a model in which both inverse dynamics control and impedance
control are active during the motor learning process to
counteract mechanical instability rather than just predicting
final learning outcomes [5]. In [6], a V-shaped learning function
is created to specify exactly how feedforward commands
are adjusted to individual muscles based on tracking error.
Tee et al. claimed the shortcomings of the aforementioned
models and defined a model specifying the activation of each
muscle that is adapted from one movement to the next and
explained how humans learn to perform movements in the
environment with novel dynamics of tool use [7].
By comparison, the model shown in [8] is based on a stability
measurement corresponding to Lyapunov exponents,
PFM
Learning
Target
DF
Start
y
(a)
[0,0]
x
+
Impedance
-
(b)
Figure 1. A simulation diagram of human movements to investigate motor control and learning (a) Point-to-point movements with
lateral instability produced by a parallel-link direct drive air-magnet floating manipulandum (PFM) to show human feedforward force and
impedance adaptation [12]. (b) A control diagram of the controller with neural control and feedback error learning in [8]. DF: divergent force.
DECEMBER 2022 * IEEE ROBOTICS & AUTOMATION MAGAZINE *
67
Planned
Motions
Feedforward
Delay
Reflex
+
+
Noise
+
+
Motor
Order
which explores the learning mechanism of human arms for
unstable interaction, and each muscle is not modeled separately
as in [7]. As shown in Figure 1, this diagram includes feedforward
forces, motor noise, delayed reflex, muscle impedance,
human body dynamics, and external forces. The feedforward
forces are determined by the planned trajectory and delayed
reflexes. The motor noise, inheriting in-motion generation, is
considered an extension of the previous model in [9]. The
impedance depends on the torque caused by muscle activation
and time delays of the reflexes. Finally, the feedforward and
feedback control units are combined to generate the muscle
force acting on the human arm to encounter external forces.
This model fully shows the human motor learning process
and greatly influences subsequent studies [10]-[16], such as
in [12], where Yang et al. proposed a sliding term to update
the feedforward torque and impedance factors (stiffness and
damping) to replace the effect of muscle contraction. Li et al.
created a controller that takes into account not only feedforward
force and impedance but also adjustment of the reference
trajectory during interactions with an unknown
environment [13]. This model is also used for impedance
control of robotic upper-limb exoskeletons [15], [16]. However,
most of the previous adaptation methods are based on iterative
learning control and adaptive control, where the
actuators' torques are updated at fixed rates (e.g., a and b in
[7]) to minimize force and position tracking errors. The rateof-force
decrease in the simulations is somehow different
from that observed in reality in [17]. Some recent learning
methods successfully used in human activities' monitoring
[18] and robot manipulation [19], [20] are helpful to recognize
and minimize the difference. The other case is that the
learning rate is different in different force fields. The experimental
results in [21] showed that absolute errors decrease
more slowly in the divergent force (DF) field than in the
velocity-dependent force (VF) field, which normally converge
exponentially within 10 trails.
For the aforementioned properties in experiments, this
article proposes a new learning and control framework based
on BLS in which the learning process is realized by increasing
the number of feature or enhancement nodes and linking
between different layers to increase or decrease the learning
Muscle System
Contractile
Component
Learning
+
+
External
Force
Muscle
Force
Body
Actual
Motions

IEEE Robotics & Automation Magazine - December 2022

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