**1. Introduction**

Exoskeletons are one of the key technologies to assist humans in a wide range of applications, such as rehabilitation, daily activities and so forth [1–3]. In particular, the adoption of exoskeletons in industrial applications is nowadays a *hot-topic* [4], since their capabilities to assist humans executing onerous tasks [5,6].

#### *1.1. Industrial Exoskeletons Design Solutions*

Industrial exoskeletons can be classified as passive and active. Passive solutions are not provided by actuation, indeed, they use springs and/or dampers to store energy from human's motion and releasing it when required [7,8]. Passive exoskeletons are commonly supporting the human operator in order to relieve him/her from repetitive tasks, while improving ergonomics [9]. Considering upper limbs solutions, different devices are available on the market [10–13]. Such exoskeletons assist humans in specific tasks, e.g., in over-the-head tasks, supporting the arm to reduce muscular stress. The main advantages of these solutions are the reduced weight and size, do not requiring motors and batteries. However, the comfortable range of postures for the human worker is restricted to specific configurations (such as in the over-the-head tasks assistance), making passive exoskeleton typically tailored to specific applications. Proposed solutions also consider a limited payload, since active assistance cannot be generated. In addition, common solutions does not support the forearm, therefore not guaranteeing the elbow support in case, e.g., of heavy parts transportation.

Active exoskeletons [14] are instead provided by actuation, allowing to empower the human worker. Different solutions have been developed in order to face different tasks, adopting different kinematics and hardware solutions [15–17]. Many solutions are available on the market, facing different tasks and scenarios. The *Panasonic Corporation* has developed an active device, called *AWN03*, that supports the operator's back when lifting heavy loads [18] thanks to electric motors. Another active human's back support has been developed in Reference [19]. *AUTON* proposes a back-exoskeleton to support workers in the transportation of heavy parts [20]. *Muscle Suit* by *Innophys* is an active back-support for lifting tasks [21]. *HAL* exoskeleton developed by *Tsukaba University* [22] is a full body exoskeleton for medical and industrial applications. *Sarcos Robotics* proposes the full-body exoskeleton *Guardian XO* for heavy tasks execution [23]. The proposed state-of-the-art solutions are still high-cost, characterized by a complex design, having a limited payload ratio (i.e., the ratio between the exoskeleton payload and the weight of the device) and not involving compliant actuation. Compliant actuation, such as series elastic actuation and variable stiffness actuation [24–26] are increasingly implemented in human-robot interaction tasks [27–29]. A compliant actuation, in fact, can intrinsically increase the safety in human-robot collaboration at the hardware level, having a deformable structure. Some upper limbs exoskeletons exploiting compliant actuation can be found for medical/rehabilitation applications [30,31]. However, due to the higher target payloads and increased design and implementation complexity, available state-of-the-art active industrial upper limbs exoskeletons are not equipped with compliant actuation. The safety issues are therefore tackled only at the software side (i.e., saturating velocities/forces). In addition, low-cost solutions can only be found in rehabilitation/medical domain, where limited power is required [32,33].

#### *1.2. Exoskeletons Control Solutions*

Exoskeleton control is widely investigated in order to assist humans in different applications [34]. Many control approaches have been developed, integrating different sensors and control techniques. Brain-control schemes have been developed exploiting a electroencephalogram signals [35]. Surface electromyograpy measurements have been exploited in order to control the exoskeleton on the basis of the human's muscles activation [36,37], also exploiting variable impedance control [38]. Admittance force control has been also exploited in order to control the exoskeleton on the basis of the measured interaction between the human and the robot [39]. External devices and measurement systems have also been used to control the exoskeleton on the basis of muscular activation, such as the Myo armband [40] or IMU sensors [41]. Common state-of-the-art approaches, however, show difficulties in the estimation of the human intention, especially while manipulating (partially) unknown payloads. Moreover, common approaches does not allow to online regulate the assistance given to the human during the task on the basis of the human-robot interaction. In addition, safety is commonly tackled in the controller only as a pre-defined saturation on the control action. Considering the empowering scenario and the manipulation of a (heavy) part, safety-based rules modulating the assistance to the

human on the basis of the current interaction state (i.e., velocity, interaction force and derivative of the interaction force) should be included in the controller.

#### *1.3. Paper Contribution*

The aim of this paper is to fill the gap above described in the industrial exoskeleton field. More in details, the paper proposes (i) the mechanical design of a low-cost exoskeleton (hardware costs <10,000 Euro) for industrial applications with (ii) high payload ration (>0.8), (iii) involving compliant actuation (to achieve intrinsic safety in human-robot interaction), together with (iv) the design of an empowering safety-based control framework.

A lifting and transportation task of a heavy component has been considered as an objective for the exoskeleton design specifications definition (case study: car bumper part with weight of 10 kg). On the basis of such task, the kinematics of the exoskeleton has been defined, together with the performance required by the exoskeleton—objective (i) and (ii). A series elastic actuator (SEA) has been designed for the shoulder joint to embed compliance into the device—objective (iii). The SEA has been designed exploiting a compliant transmission (i.e., a compliant belt) between the shoulder joint motor and the link side. The target belt compliance has been calculated in order to achieve a target equivalent shoulder joint compliance. On the basis of such specifications, components from the market (e.g., motors, elastic belt, etc.) has been selected to implement the designed solution.

The proposed empowering controller has been designed in order to actively assist the human during the task execution. Intrinsic safety rules have been embedded into the control design in order to modulate the assistance on the basis of the current interaction state (i.e., velocity, interaction force and derivative of the interaction force) —objective (iv). Furthermore, the controller has been designed to be robust to (partially) unknown payloads (i.e., the weight of the part). A hierarchic controller has been designed, composed by an inner optimal controller (for trajectory tracking purposes) and by an outer safety-based fuzzy logic controller (for human empowering purposes), online modulating the assistance. The inner model-based controller includes the compliant modeling of the shoulder joint. The outer controller (on the basis of the proposed membership functions) is capable to identify the intention of motion of the human, reacting consequently. Moreover, a gain scheduler has been designed in order to store inner optimal control gains as a function of the performed task trajectory (i.e., control gains are a function of the executed trajectory).

Simulations studies have been performed in order to validate the proposed approach, simulating different task scenarios. Simulation results are promising and the proposed methodology will be applied to the real exoskeleton for final evaluation. The proposed exoskeleton is under realization.

#### **2. Task Specifications & Exoskeleton's Design Guidelines**
