The rise of biological disasters and digital virtual ecosystems calls for the development of avatar technologies allowing humans to exist either in remote real locations or in immersive virtual realities. The recent COVID-19 pandemic, for instance, revealed the immature status of avatar technologies enabling humans to operate effectively in real remote locations . Analogously, the renewed interests of the Engineering community on virtual reality systems is also driven by the increasing applications of digital and virtual ecosystems across different segments . This drive is also exemplified by the ANA Avatar XPRIZE, a 10M$ competition111https://www.xprize.org/prizes/avatar that aims to create avatar systems transporting human presence to a remote real location in real time. Along this direction, an entire new field of applications and opportunities is emerging when approaching what is nowadays called Metaverse, namely, ecosystems where digital avatars of humans interact each for a large variety of reasons, ranging from simple virtual reality chats to creating new digital businesses . This paper contributes towards technologies and methods to create physical avatars for humans to operate effectively in remote locations via humanoid robots.
When attempting at creating physical avatars, one is tempted to apply the state of the art on telexistence , a term often interchanged with teleoperation when the latter involves advanced technologies for real time sensations from the remote site. As a consequence, a telexistence system allows transferring, and possibly augmenting, the skills of the human operator to a robotic avatar. The intuitiveness is a key feature of the system, trading off the autonomy of the robotic avatar with the capabilities of the human operator to cope with unforeseen circumstances. Through the system, the operator is physically connected in the remote location while interacting with the environment or engaging with a person.
In light of the above, physical avatar technologies benefit from the state of the art in telexistence and teleoperation. Thus, a physical avatar system is mainly composed of three components that are often the constituents of telexistence systems: the physical avatar, often a robot with a degree of locomotion; the operator system, which is in charge of retargeting and tele-perception; the communication layer, which allows the communications between the avatar and the operator system.
Physical avatars are often implemented with robots having a degree of locomotion. Typical solutions include multi-legged , or wheeled robots . In some contexts, where the remote interaction with humans is crucial, avatar humanoid robots show great potential. In fact, the human-likeness factor increases the acceptability, the social closeness to the robot, and the legibility of its intentions . At the same time, the legged design allows performing complex movements in a reduced space. Humanoid robots represent an optimal starting point for a platform able to emulate humans in terms of locomotion, manipulation and verbal and non-verbal interaction. As an example, the operator can have direct control over the whole-body posture of the robot [3, 17]. The humanoid design, on the other hand, poses additional challenges due to the inherent instability of the robotic system. This complexity can be faced by letting the robot to autonomously control its stability while achieving the desired tasks . Similarly, the operator can provide only high-level walking commands [5, 9]. In this case, the robot autonomously follows the desired walking patterns. As a consequence, the lower-body motion of the robot is not synchronized to the operator movements.
The operator system often consists of a set of wearable technologies in charge of retargeting and tele-perception. These devices are often Virtual Reality commercial products , or motion capture systems [3, 17]. In other cases, by employing special exoskeletons, it is possible to achieve bilateral feedback teleoperation. In particular, both lower-body , and full-body  exoskeletons can be adopted to fully synchronize the motion of the operator into the robotic avatar. On the other hand, these devices can be very cumbersome and invasive, constraining the motion of the operator.
The communication layer connects the operator system to the physical avatar. It allows the different components of the teleoperation system to communicate with each other, exploiting a potentially delayed network. In the robotics jargon, the software suite that implements the communication layer is referred as middleware. Common middlewares are the Robot Operating System (ROS) , and YARP .
This paper introduces an avatar system exploiting a new legged humanoid avatar: the iCub3. The robot is an evolution of the classical iCub platform, being 25 taller and 19 heavier. The operator devices are lightweight and non-invasive for better acceptability and immersion. Moreover, we exploit both off-the-shelf and custom-made devices. We demonstrate the teleoperation system by allowing a human operator in IIT, Genoa, to remotely control the iCub 3 humanoid robot in the Italian Pavilion within the Biennale dell’Architettura di Venezia, Venice, at about 290 distance. Using a standard fiber optic internet connection, the avatar system transports simultaneously the operator locomotion, manipulation, voice, and face expressions to the robotic avatar with visual, auditory, haptic and touch feedback. To the best of the authors’ knowledge, it is the first time that such a complete avatar system is validated on a legged humanoid robot allowing immersive and remote verbal, non verbal and physical interaction.
Ii The Avatar System Architecture
The avatar architecture presented in this paper is depicted in Fig. 2. The figure describes both the physical and the logical connections between the operator and the avatar. The former is achieved by a network of computers operating on both the operator and robot side, connected via a standard fiber optic internet connection.
The physical network allows achieving a set of logical connections between the operator and the robot. The operator’s actions, intentions and emotions are captured by a set of devices the operator is wearing. These are part of the retargeting component and are transmitted in the form of references to the avatar control. The retargeting and control compose the teleoperation interface.
The second interface is the teleperception. The measurements retrieved by the robot are transmitted to the operator as a feedback, providing a first-person perspective of the surroundings sensed by the robot.
The following presents the iCub3 avatar architecture. First, we present the avatar, iCub3, in Sec. III. Then, we introduce the communication layer in Sec. IV, connecting iCub3 to the operator system, presented in Sec. V. In this last section, we present the set of devices used by the operator that allow teleoperation and teleperception.
Iii The Avatar: iCub3
The longstanding iCub platform has been evolving along several directions over the last fifteen years . However, all its versions222See also https://icub-tech-iit.github.io/documentation/icub_versions for the precise details on the iCub versions., which range from v1.0 to v2.9, have concerned a humanoid robot having mostly the same morphology, size, joint topology, actuation and transmission mechanisms. In other words, the evolution of iCub mechanics never concerned the robot height – which kept being of about one meter – nor the robot actuation and transmission mechanisms – which never evolved for the robot to increase its dynamism substantially – nor its force sensing capabilities – which are derived from Force/Torque sensors of 45 mm diameter installed in the robot . The iCub3 humanoid robot shown in Fig. 1 is the outcome of a design effort that takes a step in all these directions. The robot represents a concept of humanoid that will be the starting point when conceptualising the next generations of the iCub platform.
The iCub3 humanoid robot is 125 tall, and weighs 52. Its mechanical structure is mainly composed by an aluminum alloy. The robot also presents plastic covers that partially cover the electronics. The weight is distributed as follows: about 45% of the weight is on the legs, 20% on the arms, and 35% on the torso and head.
Each robot leg is approximately 63 long, while the arms are 56 long from the shoulder to the fingertips. With the arms along the body, the robot is 43 wide. Each foot, showed in Fig. 3, is composed of two separate rectangular sections, with a total length of about 25 and 10cm wide.
The iCub3 robot possesses in total 54 degrees of freedom including those in the hands and in the eyes, and they are all used in the avatar system. They are distributed as follows:
4 joints in the head, controlling the eyelids and the eyes,
3 joints in the neck,
7 joints in each arm,
9 joints in each hand,
3 joints in the torso,
6 joints in each leg.
The iCub3 hands are equipped with tendon driven joints, moved by 9 motors, allowing to control separately the thumb, the index and the middle finger, while the ring and the pinkie fingers move jointly .
The iCub3 is equipped with both DC and brushless three-phase motors.
The DC motors actuate the joints controlling the eyes, the eyelids, the neck, the wrists and the hands. They are equipped with a Harmonic Drive gearbox with 1/100 reduction ratio. The reduced dimensions of these motors make them suitable for controlling the joints that do not require high torques.
The torso, the arms and the legs are controlled by three-phase brushless motors, also coupled with 1/100 Harmonic Drive gearbox. The motor charachteristics are as follows. The rated power is 110, with a rated torque of 0.18, while the continuous stall torque is 0.22.
The hip pitch, knee, and ankle pitch joints are driven by another type of brushless motor, bigger and more powerful. For this set of joints, the rated power is 179, with a rated torque of 0.43 and a continuous stall torque of 0.48.
The iCub3 robot is powered either by an external supplier or by a custom made battery. The connection to the robot can be established through an Ethernet cable or wirelessly via a standard 5GHz Wi-Fi network.
The robot head is equipped with a 4 generation Intel® Core email@example.com computer with 8 of RAM and running Ubuntu. This central unit represents the interface between the robot and the other laptops in the robot network, Fig. 2.
The iCub3 central unit is communicating with a series of boards distributed on the robot body and connected via an Ethernet bus, as illustrated in Fig. 4. There are two main types of boards connected to the bus:
the Ethernet Motor Supervisor (EMS) boards, controlling the three phase motors;
the MC4Plus boards, controlling the DC motors.
The EMS board is a 32-bit Arm Cortex micro-controller. It runs at 1 and implements different control strategies, described in Sec. III-F. It communicates via CAN protocol with the motor driver board (2FOC), which generates PWM signals at 20 KHz to drive the motor.
Similar to the EMS board, the MC4Plus board is a 32-bit Arm Cortex micro-controller implementing different motor control strategies, running at 1. Differently from the EMS though, the MC4Plus board can control directly up to four DC motors. They control the neck, wrists and hands joints.
A particular feature of iCub3 is the vast array of sensors available. More in detail, iCub possesses 6 six-axes force/torque (F/T) sensors . Two of them are mounted at the shoulders, and two on each foot, connecting the two sections of the feet to the ankle assembly.
iCub3 also possesses tactile sensors as an artificial skin  on the upper arm and the hands, which provides information about both the location and the intensity of the contact forces.
The head sports several sensors. It possesses two cameras capturing images at 15 frames per second, with a resolution of 1024x768. The cameras are placed within the eyes bulb and can be controlled to a specified vergence, version and tilt angle. Both eyes are equipped with eyelids, controlled jointly by a single DC motor. The robot head also includes a microphone on both hears, and a speaker behind the face cover. Finally, a set of LEDs define the robot face expression.
At the joint level, the iCub3 robot uses a series of encoders. The motors controlled by the 2FOC board need of incremental optical encoder mounted on the motor axis. These allow to estimate the motor magnetic flux. At the same time, the EMS boards exploit an off-axis absolute magnetic encoder mounted on each joint, after the gearbox, to estimate each joint position and velocity.
In the robot hands, the MC4Plus boards communicate with a series of MAIS boards. These are 32-channel miniature ADC boards used to acquire the measurements from the skin and from the hall sensors installed in the robot fingers, used to estimate the position of each phalanx.
Iii-E Comparison with the classical iCub platform
With respect to a classical iCub platform , the iCub3 humanoid robot is bigger, being 25 taller, and weighing 19 more. Figure 5 shows the different dimensions of the two platforms. The increased weight requires more powerful motors on the legs. As a result, the increased dimension of the actuators required a different approach for the knee and ankle pitch joint. In particular, instead of having the motor and the actuator on the same axis, they are displaced and connected by belts. Moreover, the torso and shoulder joints are serial direct mechanisms, while classical iCub robots have coupled tendon-driven mechanisms. This allows higher range of motion and greater mechanical robustness.
In addition, iCub3 has a higher capacity battery, 10050mAh versus 9300mAh, and this is part of the torso assembly instead of being included in a rigidly attached backpack.
The head of iCub3 is identical to a classical iCub, although the neck is longer for better proportions. Similarly, also the hands are in common between the two versions. The only difference is given by the location of the electronics board that control the finger. In the iCub 2.7, they are located in the upper arm, while on iCub3 they are housed in the forearms.
From the electronics point of view, both platforms share the same 2FOC/EMS/MC4Plus architecture, although iCub3 has higher resolution encoders. On the sensors side, the iCub3 platform has also an additional Intel Realsense D435i depth camera, while the eye cameras have better resolution compared to the classical iCub.
Iii-F Low Level Robot Control
Both the EMS and MC4Plus boards, described in Sec. III-C, can control the robot joints using position and velocity mode. Moreover, the EMS board also support controlling the joints in torque. We briefly present here the different control modes:
the position control mode implements a PID controller that aims to track a trajectory generated imposing a minimum jerk profile. A variant of the position control mode is called position direct and differs from the previous one as it does not impose any constraint on the acceleration. Both these control modes close the loop on the measured joint position.
The velocity control implementation includes an integration of the velocity reference to compute the desired joint position tracked then, using the position control mode.
The torque control mode uses a PID controller closing the loop on the joint torque estimated using the F/T sensor readings . The PID controller output is combined with a feed-forward and a mechanical friction compensation.
The 2FOC board implements two different control modes:
The current control mode. It implements a PI controller closing the loop on the measured current. The controller output is transformed into a PWM used to drive the motor.
The PWM control mode. This does not consider any internal loop and the motor is driven in open-loop.
Iii-G High Level Robot control
The robot motion is controlled adopting a layered control architecture, described in .
Each loop of the architecture receives inputs from the robot and the environment, and provides references to the loop next. We refer the reader to Fig.2 of  for a schematic representation of the architecture. The inner the layer, the shorter the time horizon that is used to evaluate the outputs. Also, inner loops employ more complex robot models to evaluate their outputs. More precisely, from outer to inner, the hierarchical control architecture is composed of the following loops: the trajectory optimization; the simplified model control; and the whole-body quadratic programming (QP) control loop.
The trajectory optimization loop is in charge of generating foothold trajectories from high-level commands, like the desired walking direction and speed. This layer mainly deals with defining the contact locations of the locomotion pattern.
The output of the trajectory optimization layer is given to the simplified model control loop. Its aim is to generate desired and feasible centroidal quantities  associated with stable walking instances. This loop exploits the contact force feedback, measured thanks to the force/torque sensors installed on the robot feet.
Finally, the whole-body QP control layer, which is in charge of stabilizing the planned trajectories exploiting the full robot model using a suitable Quadratic Programming (QP) formulation. This layer uses the dynamic model of the system to ensure the tracking of the desired trajectories. The control problem is formulated using the stack of tasks approach. The tasks are divided into high and low priority. The high priority tasks are considered constraints of the QP problem and are designed to track the trajectories for the center of mass, feet, and root link height generated by the simplified model control layer. The low priority tasks are designed to keep the torso in a desired orientation, while controlling the joint posture around a specified reference. This joint level reference is obtained by means of geometric retargeting . In other words, the posture of the operator is directly used as a reference for the joint regularization.
The output of the whole-body QP control layer is a joint position reference. This is tracked by the PID controllers running on the iCub motor control boards.
Iv The Communication layer
Both the robot and the operator system require a cluster of different PCs connected in a local area network (LAN), running multiple applications at once on different operating system. The communication between the different applications is done through YARP .
YARP supports building a robot control system as a collection of programs communicating in a peer-to-peer way, with an extensible family of connection types, like TCP, UDP, or other carriers tailored for the streaming of images.
For real-time operation, network overhead has to be minimized, so YARP is designed to operate on an isolated network or behind a firewall. On the other hand, the operator and the robot might be on two different far places. In order to have the two sub-networks connected to each other, we use OpenVPN333https://openvpn.net/, so that YARP can still be used to have communication between applications running on either side of the network.
A simplified diagram of the robot and operator network is depicted in Fig. 2. Each network possess a simple Ethernet switch that allows connecting multiple PCs, and the robot (either via wired or wireless connections). Each local sub-network is controlled by a router that is also connected to Internet. Both the two routers connect to the same OpenVPN server, thus joining in a Virtual Private Network (VPN). Once the two routers are connected, all the devices connected to both switches are visible to each other transparently.
V The Operator system
In the iCub3 avatar system, presented in Fig. 2, the operator is exploiting a series of devices:
the HTC Vive PRO eye444https://www.vive.com/eu/product/vive-pro-eye/overview/ headset,
the VIVE Facial Tracker555https://www.vive.com/eu/accessory/facial-tracker/
the iFeel sensorized and haptic suit666https://ifeeltech.eu/
the SenseGlove DK1777https://www.senseglove.com/ haptic glove
the Cyberith Virtualizer Elite 2888https://www.cyberith.com/virtualizer-elite/ omnidirectional treadmill.
The operator devices define the retargeting and feedback interfaces defined in Fig. 2.
V-a The retargeting interfaces
The retargeting interfaces contain the set of commands that the operator exploits (on the robot) to achieve a specified task in the remote environment. In the iCub3 avatar system, we can distinguish the following retargeting interfaces:
The manipulation interfaces exploit the headset, sensorized suit and glove to control respectively the robot head, body, and fingers. In particular, the 3D pose of the headset is tracked in real-time, defining a reference for the motion of the robot head. The operator gaze and eye openness are tracked too, allowing to retarget directly the robot eyelids and gaze. The robot upper-body motion is controlled via the iFeel sensorized suit. The suit exploits an array of devices, called iFeel nodes.
The iFeel nodes contain an integrated inertial measurement unit (IMU) and instantiate a wireless connection with a base station connected to a PC. Each iFeel node estimates the pose of the operator limb to which they are attached. The core strengths of these devices are the reduced weight, they are non-invasive, and the system is highly flexible. In particular, we exploit five nodes to estimate the orientation of the operator arms with respect to his chest, exploiting the geometric retargeting approach presented in . Through the geometric retargeting approach we compute a set of desired joint values to be fed to the robot controller presented in Sec. III-G.
The SenseGlove haptic glove completes the set of devices of the manipulation interface. It is an exoskeleton-like haptic glove allowing to translate the motion of each of the operator’s fingers into a reference for the robot fingers.
The locomotion interface is supported via the Cyberith Virtualizer Elite 2. It is an omni-directional treadmill where the operator walks by sliding. The motion is detected through optical sensors located on the device base plate, where the operator is walking. The motion direction is estimated via a moving ring attached to the harness secured to the operator waist. The base plate can also be inclined of a fixed amount to ease the sliding motion, allowing the operator to walk naturally. The walking motion of the operator is not directly commanded to the robot, but rather interpreted as a reference walking direction and speed . These references are fed to the planning layer presented in Sec. III-G.
Finally, the voice and face expressions interface exploit the HTC VIVE headset microphone and the attached VIVE facial tracker. The former allows the operator to verbally interact through the operator. The latter is fundamental for the non-verbal interaction. One of the particular features of the headset and the facial tracker is the capability of estimating the operator’s face expressions. These are replayed by the robot LEDs installed on its face.
V-B The feedback interfaces
The feedback interfaces report the robot sensors measurement to the operator. In the iCub3 teleoperation system we have the following feebdack interfaces:
The headset is fundamental for the visual and auditory feedback. The images captured by the robot cameras are displayed inside the headset, allowing the operator to have a first person view of what the robot is seeing. At the same time, the audio captured by the robot microphones is directly played on the headset’s headphones.
The robot sensorized skin and the iFeel haptic nodes are fundamental for the body haptic feedback. A touch on the robot arm is mimicked by a vibration on the robot arms, thus providing a mean of physical interaction.
Finally, the SenseGlove provides touch feedback by means of vibration motors in each fingertip and through a set of brakes able to produce up to 20 of passive force per finger.
The iCub3 avatar system is tested in a demonstration involving the operator located in the IIT offices in Genova, Italy, and the iCub3 robot in the Italian Pavilion of the Biennale dell’Architettura located in Venice, Italy. Hence, the operator and the robot are about 290 apart, “connected” via a standard fiber optic internet connection. The test has been performed on November the 8th, 2021. The latency introduced by the communication channel has been constantly monitored, remaining stably below 25. This reduced latency did not affect the operator experience. In addition, this delay does not affect the robot stability since it is able to keep the balance independently from the network configuration. The video of the demonstration is available at https://youtu.be/r6bFwUPStOA.
The first part of the video, up to 0:55 is dedicated to the preparation of the operator, who is wearing the devices mentioned in Sec. V. The operator is then virtually transported to the remote location.
At 1:25, and later at 1:51, the operator exploits the robot locomotion capabilities. In particular, by walking inside the Cyberith Virtualizer platform, the operator is able to walk around the venue, as showed in Fig. 11. At 1:26, the operator then interacts through the avatar with a person in the remote location, here defined as recipient. In this context, the visual and auditory feedback are fundamental for a proficient verbal interaction. The face expressions retargeting, demonstrated in Fig. 8, enables the non-verbal interaction, allowing the operator to smile to the recipient, or to close the eyes in case of a strong light, as demonstrated in Fig. 12, and in minute 2:19 of the accompanying video.
In minute 1:58 and 2:08 the operator exploits the control over the robot body to express body language and pointing to some installations while interacting with the recipient.
The touch feedback is fundamental when the operator interacts with the venue at 2:43, Fig. 13. The manipulation and fine control of each robot finger allows the operator to touch the installation with delicacy, while perceiving haptic feedback.
Finally, at 2:52, we showcase the importance of the body haptic feedback for an immersive interaction. As showed in Fig. 18, the recipient reaches the robot from outside its field of view. She then touches the robot arms. The robot skin perceives the touch and triggers the body haptic feedback. Hence, the operator perceives the remote touch and turns toward the recipient direction.
The remote visit ends with the operator and the recipient sharing a hug, highlighting the emotional implications of such rich interaction.
We present a demonstration where an operator teleoperates the humanoid robot iCub3 to visit a remote exhibition. The operator is able to walk around the remote location, while interacting physically, verbally and non-verbally with another person through the robotic avatar.
This demonstration was possible thanks to a collaboration between IIT and the Italian Ministry of Culture. Human avatars via humanoid robots may have tremendous applications, ranging from disaster response, telepresence, and teletourism.
As a future work, we plan to increase the overall robustness of the robot, so that it can endure and recover from falls, for example. In addition, we plan to improve the overall autonomy and usability of the system allowing severely disabled people to operate in remote locations via the robotic avatar.
-  (2018) Humanoid teleoperation using task-relevant haptic feedback. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 5010–5017. Cited by: §I.
-  (2008-08) An embedded artificial skin for humanoid robots. In Multisensor Fusion and Integration for Intelligent Systems, 2008. MFI 2008. IEEE International Conference on, pp. 434–438. Cited by: §III-D.
-  (2019) Whole-body geometric retargeting for humanoid robots. In 2019 IEEE-RAS 19th International Conference on Humanoid Robots (Humanoids), pp. 679–686. Cited by: §I, §I, §III-G, §V-A.
-  (2013) Legibility and predictability of robot motion. In 2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pp. 301–308. Cited by: §I.
-  (2019) Telexistence and teleoperation for walking humanoid robots. In Proceedings of SAI Intelligent Systems Conference, pp. 1106–1121. Cited by: §I, §V-A.
-  (2014) A middle way for robotics middleware. J. Softw. Eng. Robot 5 (2), pp. 42–49. Cited by: §I, §IV.
-  (2012) Force feedback exploiting tactile and proximal force/torque sensing. Autonomous Robots 33 (4), pp. 381–398. External Links: Cited by: §III-D, §III.
-  (2020) Bilateral humanoid teleoperation system using whole-body exoskeleton cockpit tablis. IEEE Robotics and Automation Letters 5 (4), pp. 6419–6426. Cited by: §I.
-  (2013) Whole body motion control framework for arbitrarily and simultaneously assigned upper-body tasks and walking motion. In Modeling, Simulation and Optimization of Bipedal Walking, pp. 87–98. Cited by: §I.
-  (2020) Remote mobile manipulation with the centauro robot: full-body telepresence and autonomous operator assistance. Journal of Field Robotics 37 (5), pp. 889–919. Cited by: §I.
-  (2021) The covid-19 pandemic: an accelerator for the robotics industry?. IEEE Robotics & Automation Magazine 28 (1), pp. 116–116. Cited by: §I.
-  (2015) Telepresence: immersion with the icub humanoid robot and the oculus rift. In Conference on Biomimetic and Biohybrid Systems, pp. 461–464. Cited by: §I.
-  (2017) ICub: the not-yet-finished story of building a robot child. Science Robotics 2 (13). Cited by: §III-E, §III.
-  (2015) ICub whole-body control through force regulation on rigid non-coplanar contacts. Frontiers in Robotics and AI 2, pp. 6. Cited by: 3rd item.
-  (2008) Centroidal momentum matrix of a humanoid robot: structure and properties. Intelligent Robots and Systems, 2008. IROS 2008. IEEE/RSJ International Conference on, pp. 653 – 659. Cited by: §III-G.
-  (2022) A metaverse: taxonomy, components, applications, and open challenges. IEEE Access. Cited by: §I.
-  (2019) A multimode teleoperation framework for humanoid loco-manipulation: an application for the icub robot. IEEE Robotics & Automation Magazine 26 (4), pp. 73–82. Cited by: §I, §I.
-  (2022) Extended reality usage during covid 19 pandemic. Springer. External Links: Cited by: §I.
ROS: an open-source robot operating system. In ICRA workshop on open source software, Vol. 3, pp. 5. Cited by: §I.
-  (2018) Humanoid dynamic synchronization through whole-body bilateral feedback teleoperation. IEEE Transactions on Robotics 34 (4), pp. 953–965. Cited by: §I.
-  (2018) A benchmarking of dcm based architectures for position and velocity controlled walking of humanoid robots. In 2018 IEEE-RAS 18th International Conference on Humanoid Robots (Humanoids), pp. 1–9. Cited by: §III-G, §III-G.
-  (2010) Design, realization and sensorization of the dexterous icub hand. In 2010 10th IEEE-RAS International Conference on Humanoid Robots, pp. 186–191. Cited by: §III-A.
-  (2021) NimbRo avatar: interactive immersive telepresence with force-feedback telemanipulation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 5312–5319. Cited by: §I.
-  (2015) Telexistence. In Virtual Realities, pp. 229–259. Cited by: §I.