In this chapter, we will first give a brief review of the state-of-art developments in sEMG-based stiffness transfer from humans to robots. Subsequently, in Section 1.2, we will introduce several commonly used approaches that can model human skills and thus enable the skill transfer to robots. In Section 1.3, the design of some intelligent controllers that can enhance robotic skill learning will be introduced, such as admittance control, variable impedance control, and neural network-based controllers.
1.1 Overview of sEMG-Based Stiffness Transfer
It is predicted that humans and robots will be closely working together, sharing workspace, and thus collaborating to fulfil sophisticated tasks in the near future. The experimental results have demonstrated that a human-robot team will be more efficient and flexible than human or robot working alone [1, 2, 3]. Transferring skills from human tutor to robots, especially human impedance adaptive skills [4], is seen as one of the most effective ways to improve the efficiencies of human-robot collaborative systems. Skill transfer is generally defined as the act of learning manipulative skills for the robot according to the demonstration by the human tutor [5, 6].
Research of human motor behavior reveals that human arm can be stabilized mainly using mechanical impedance control during interaction with a dynamic environment [7] which minimizes the interaction force and performance errors. Since motion trajectory transfer could not allow a robot to generate desired stiffness, e.g., the robot manipulator is unable to operate in a compliant or a rigid manner. Therefore it is difficult for the tutor to guarantee compliance, safety, and efficiency of robot manipulation due to variation of tasks and environmental uncertainty. In contrast, impedance control could enable a target relation between force and displacement. It is essential for robots that come in contact with human or the environment to operate in a safe and natural way. The work [7] reported that human central neural system (CNS) is able to adapt endpoint impedance voluntarily when human performs a task. It is thus of great importance to transfer human tutor's adaptive impedance to a robot in dynamic environments to enable human-like compliance and adaptability. A robot could have a better performance by learning human impedance featured skills, such as calligraphic writing [8], welding [9], and switching [10]. Inspired by these research results, biomimetic learning controllers are proposed in Refs. [11, 12, 13] which are able to simultaneously adapt force, impedance, and trajectory in the presence of unknown dynamics. Compared to traditional robotic controllers, they are “human-like” which enabling robots to have some human motor features in an economic perspective and therefore may have great potentials in compliant human-robot interactions, especially in some scenarios with physical contact, e.g., rehabilitation or daily tasks.
There exist different kinds of technologies for transferring human mechanical motion, impedance, or motor control mechanism to robot, e.g., various body sensors and mathematical models. A vision-based model may be a good candidate in transferring human limb movements to the robots [14, 15], but could not transfer force or impedance to the robot such that the transparency among human, robots, and environment may be attenuated. Alternatively, sEMG signals may be ideal bio-signals to incorporate human skills into robots. They reflect human muscles activations that represent human joint motion, force, stiffness etc. [16, 17, 18, 19]. Moreover, sEMG signals are easily accessible and fast adaptive, used in different applications (e.g., rehabilitation, exo-skeleton) coupled with force, sound, or vision sensors [20, 21, 22, 23, 24], etc. Therefore, sEMG signals are widely used for robots to understand human motion intention during implementing tasks.
Generally, sEMG signals can be processed into two divisions: finite class recognition serials and continuous control reference. The former usually refers to pattern recognition, such as hand posture recognition [25, 26] and such data serials are usually used as switch control signals; in contrast the latter refers to extract continuous force, stiffness and even motion serials from sEMG signals which reflect the variations of human limb kinematics and dynamics during limb movement or pose maintenance. Furthermore, the relationship between sEMG and stiffness, force and motion is approximately linear [16], and thus bio-controller design tends to be simple in sEMG-based robot control system. In Ref. [27], sEMG signals are processed to extract incremental stiffness to reduce stiffness estimation error and calibration time. Its application was tested via robot anti-disturbance pose maintenance. In Ref. [28], tele-impedance is implemented via continuous stiffness reference, and in Refs. [29] and [30], finger position from sEMG signals are continuously estimated though with rather relatively large error. In Ref. [31], squaring and low-pass filtering-based signal envelop extraction algorithm, as well as re-sampling method, is employed to extract incremental smooth stiffness from sEMG signals which are then transferred to the robot to mimic human motor behaviors.
As far as humanoid robot manipulator is concerned, it is ideal for transferring human limb dynamic features to the manipulator with elastic actuators because of their geographical similarity inspired by Ref. [13]. There will be many advantages of this human-robot dynamic transfer such as safety, compliant interaction with human, and environment with low contact force, small trajectory errors, and less time consumption [32]. In Ref. [31], sEMG-based writing skill transfer is proposed. Continuous incremental joint stiffness is extracted from sEMG signals during arm implementing tasks and then transferred to the robot arm via a mapping mechanism under robot stable boundaries.
The interactive interface is usually considered as a bridge between the human tutor and a robot for skill transfer. It plays a great role, especially in the scenarios that require impedance regulation skills [33]. Therefore, the interface needs to be carefully considered [34, 35, 36, 37]. Several types of interfaces have been developed to reach this goal. Conventional interfaces such as keyboard, joysticks, or human motion capture device such as Leap Motion, are implemented in most of the simple tasks through programming. But they are not applicable or suitable to complicated interactions involving the sensorimotor feedback. Some emerging techniques, such as virtual reality [38] and augmented reality [39], have been recently introduced into pHRI systems as communication interfaces. Generally, such interfaces with sophisticated controllers and sensors are efficient in position control by providing human tutor virtual sensor feedback. However, only position control with virtual feedback is not sufficient for a robot to fulfil complex and flexible tasks. Interactive forces need to be carefully regulated as well in the scenarios where human and robot will inevitably touch each other physically [40, 41].
More recently, the new generation of robot platforms such as the Baxter robot can be taught by using ...