Enabling Haptic-Integrated Interactive Holographic Video Streaming Powered by 5G Edge Computing

Enabling Haptic-Integrated Interactive Holographic Video Streaming Powered by 5G Edge Computing Peng Qian∗, Ning Wang†, Carl Udora‡, Carlos Velez Redondo‡, JingXuan Men‡, Rahim Tafazolli‡ ∗Department of Engineering Science, University of Oxford, Oxford, U.K. †School of Electrical, Electronic and Mechanical Engineering, University of Bristol, Bristol, U.K. ‡5GIC & 6GIC, Institute for Communication Systems (ICS), University of Surrey, Guildford, Surrey, U.K. Emails: [email protected], [email protected], {cu00029, cv00218, j.men, r.tafazolli}@surrey.ac.uk Abstract—Driven by the rapid advancement of XR technologies, an increasing demand has emerged for enabling interactive holographic real-time applications of multiple users. However, in wireless systems, the challenges of practically delivering a satisfactory experience for such applications arise not only from the high bandwidth and computational demands of holographic content but also from the need to effectively integrate and transmit additional sensory data, such as haptic feedback, to convey accurate and human-perceivable semantics. In this paper, we propose an architecture that leverages the computational capabilities of 5G edge computing to collaborate with multi-user haptic-integrated holographic video applications. We detail the components of this solution architecture, including haptic-integrated holographic video capture and streaming, and a simple reinforcement learning algorithm to synchronise multiuser frames at the MEC server. A prototype of a representative hand-touch social interaction application was designed, implemented, and measured to exemplify the potential of human-perceivable, haptic-integrated holographic video applications. Index Terms—Holographic Teleportation, extended reality (XR), 5G, edge computing, haptic-integrated holographic video streaming I. INTRODUCTION In recent years, innovations in extended reality (XR) devices have driven research to enhance the user experience in holographic applications. This rapid development has catalysed diverse applications, including remote healthcare, gaming, education, and industrial operation. However, most current applications are based on a single sensory modality. Integrating diverse sensory inputs (e.g. haptic data) into holographic video streaming has thus become a key feature of real-time holographic technologies. Meanwhile, the advent of 5G networks raises the critical question of whether reliable and perceptibly interactive experiences are feasible in mobile environments [1], [15]. This progress presents two key challenges: first, how user-side hardware can efficiently capture holographic video in real time, transmit it via the uplink, and simultaneously render remote holographic data; and second, how to ensure social interactions are properly perceived through synchronised visual and tactile experiences, between multiple users. Addressing these challenges requires seamlessly integrating haptic feedback with holographic video through a robust network architecture and protocol design, including effective hardware integration. These challenges are multifaceted and stem primarily from the following aspects: First, holographic video based on point cloud technology not only demands high transmission rates but also requires robust computational capability for continuous and accurate point cloud processing (e.g. calibration, completion, and rendering). Second, the primary functions of data capture devices focus on the acquisition and triggering of sensory data, lacking a unified protocol standard or software solution to integrate these devices and ensure reliable transmission over wireless networks. Third, accurate conveyance of semantic information composed of visual and haptic frames requires that the latency of both types of frames be delivered within a given temporal threshold [6]. For instance, when observing the touch of an object via the holographic vision in the headset, the corresponding force feedback must be promptly reflected in the haptic gloves to the observer. Considering the challenges described, we have proposed a solution architecture to support haptic-integrated holographic video streaming use cases in 5G networks. The key component is the Mobile Edge Computing (MEC) Server. Firstly, its computational resources not only enable remote production services for data collected by mobile terminals, but also apply enhancement strategies such as compression, calibration, and completion. Secondly, the MEC server monitors multi-user and multi-link conditions, serving as a central point for aggregating, synchronising, and distributing haptic-integrated holographic video streaming data. Third, a reinforcement learning (RL) algorithm is employed to synchronise frames across multiple users, ensuring a seamless and real-time interactive experience. Based on this, we have designed a “remote hand touch with force feedback” scenario to demonstrate real-time hand touch experiences with reduced latency. To implement and evaluate the proposed framework, we outline a solution architecture using advanced hapticintegrated holographic video equipment and a real 5G network. Experimental results demonstrate that interactive activities like hand touch can be supported at low resolution, offering insights for future haptic-integrated holographic applications. The RL algorithm tests also show its effectiveness in minimising perception time differences between users, ensuring a synchronised and real-time experience. II. RELATED WORKS In this section, we provide a comprehensive review of the literature on recent advances in the transmission of holographic and tactile content. A. Real-time Holographic Applications and use cases The primary challenge of live transmission of point-cloud videos is the large bandwidth required. Toward a reduced traffic volume while retaining satisfactory user experience, existing methods can be categorised into two directions. An approach involves using machine learning techniques to learn the features of the point cloud data to be transmitted, thereby only sending these smaller feature information and reconstructing the original point cloud content at the receiver’s end. A representative work is [2], where the authors leverage machine learning techniques to split a large frame into different subframes of different viewpoints and carry out customised selection according to the viewpoint of the viewer. The results indicate that the edge server is capable of boosting QoE by up to 85%. However, a common drawback of these works is that machine learning-based algorithms for encoding and decoding the features of the point cloud require significant computation time [3], making them impractical for live transmission. On the other hand, there are also optimisations by retaining general compression ratio techniques but improving the transmission architecture and the point cloud transmission process [15]. For example, the authors in [15] offload the time-consuming point cloud computation task to the MEC server and examine the effect of different congestion control algorithms and resolution levels, providing an efficient solution to deploy holographic application in practical mobile network. Similarly, the authors in [11] propose a CoMIC collaboration framework that aims to address the lack of infrastructure that supports multiuser XR applications by creating a collaborative, visual first and hologram-based computing space. These enhancements continuously reduce the bandwidth needed for realtime point-cloud transmission. By admitting that both of these solutions are still in their infancy stage, academics have initially established a fundamental framework for semantic communication, defining its objectives, semantic encoding and decoding mechanisms, and transmission models [16]. By abstracting captured objects and scenes into semantic information, the receiver can even leverage generative artificial intelligence to reproduce the content from such semantic information (e.g. text). B. Tactile network The tactile network, essential for transmitting haptic information in real-time applications such as remote surgeries and industrial automation, is still in its development stages compared to holographic video frameworks. The authors in [4] highlight the critical role of 5G’s low latency in the robotic control use cases, emphasising the necessity for advanced network management to meet the stringent demands of next-generation robotic applications. The authors in [5] discuss the integration of edge computing with the tactile Internet to support human digital twin (HDT) applications. Details the system architecture required for HDT, focussing on real-time feedback, high-fidelity virtual modelling, and ultra-high reliability. The paper also covers significant design challenges, such as ensuring widespread connectivity and real-time interaction, with a case study on physical therapy that shows the effectiveness of the proposed framework. The authors in [6] investigate the perceptibility of visualhaptic desynchronisation in virtual environments. It finds that participants do not notice this desynchronisation less than 50 ms after visual contact, with only a 15 ms delay being tolerable for haptic feedback before visual stimuli. This research offers crucial insights for Virtual Reality (VR) design, stressing the need to minimise latency to boost the realism and effectiveness of haptic-integrated holographic video streaming interactions. The authors in [7] introduce a modular testbed designed for research on Tactile InternetBased Cyber-Physical Systems (TCPS). It integrates various components to allow testing of different hardware, algorithms, and communication protocols, focussing on latency analysis and performance enhancement under suboptimal network conditions. Based on reviews of holographic and tactile works, there has not yet been any work that combined them over 5G networks for human-to-human interaction applications. This gap necessitates our proposed experimental framework, which must consider the uncertainties of 5G links and the significant transmission and processing time discrepancies among haptic-integrated holographic video streaming data. III. FRAMEWORK DESIGN In this section, we present a live social interaction use case with holographic and haptic feedback as an example to illustrate how the proposed framework orchestrates its network and sensor components to cohesively enable multisensory communication (see Fig. 1). Furthermore, a pluggable Qlearning (QL) algorithm [8] is designed and explained to optimise frame synchronisation between multiple users with a focus on its adaptability in handling dynamic real-time scenarios. A. Live social interaction with haptic force feedback The proposed framework provides a shared virtual environment in which two users, located in different physical locations, can engage in interactive social activities such as virtual handshakes and collaborative object manipulation. Holographic imaging, along with the user’s real-time location, hand posture, and finger information, is captured and digitised into frames in real-time. These frames are continuously transmitted through the 5G uplink to a mobile edge server, where they are processed and forwarded through the 5G downlink to the counterpart user. The data is then processed and decomposed into formats required by various sensory devices, triggering corresponding feedback and forming information with practical semantics (e.g. hand touches). Each type of sensory information that can be captured and triggered on the user side is described below. Holographic video: The volumetric representation of a user’s hologram is built upon dense point cloud data, where each point integrates both RGB colour and depth information, typically acquired through sensors such as the Microsoft Azure Kinect DK [14]. These point clouds are processed and rendered in real time on XR head-mounted displays, with rendering latencies under 40 ms for low-resolution Fig. 1: Proposed haptic-integrated holographic video streaming architecture with 5G edge computing scenarios [14]. Each point is defined in a three-dimensional Cartesian space (X, Y, Z), thus allowing full six-degree-offreedom (6-DoF) spatial tracking. Due to the sheer volume of data, especially in high-fidelity holographic streams, the bandwidth demand can reach several hundred megabytes per second, or even approach gigabit levels. Despite applying compression techniques, transmission over the 5G uplink still suffers from noticeable performance variability [15]. User position: Real-time localisation data can be captured using local signal emitters and terminal trackers, such as HTC VIVE Trackers [12] and Steam Base Stations. By tracking within a defined 3D region (e.g. 2m3), the system offers fine-grained spatial positioning at millimetre level and updates at millisecond intervals, ensuring highly accurate virtual environment mapping. The spatial positions of the tracked body parts are represented in X, Y, Z format. With multiple trackers and at least two base stations, the system achieves a retrieval interval of 22 ms and a measurement error at the millimetre level. Haptic data: Advanced haptic gloves such as the SenseGlove Nova [9] capture intricate metrics of the hand joint, such as bending angles and range of motion, and use electromagnetic braking systems to simulate force feedback that reaches 20 N, allowing for lifelike tactile interactions. This vibration feedback further enhances tactile realism, improving interaction with virtual objects or other users. Data integration and transmission design: On the user side, the above mentioned sensory data types are captured, integrated via a client side programme, and then transmitted to the MEC server. All data streams are encapsulated into structured binary formats and delivered using the TCP protocol to guarantee ordered and reliable transmission across the network. To ensure accurate matching across modalities, each sensory data type is assigned a distinct frame identifier along with a synchronised timestamp, enabling their proper association during transmission and processing. On the MEC server side, upon receiving these sensory data frames from multiple users, it first aligns and synchronises the holographic video, haptic, and localisation data using timestamps and frame IDs. The MEC server processes each type of data as follows: The visual frames are calibrated with positional and hand movement information, projecting multi-user’s data into a unified virtual space. Within this space, the server detects user actions and postures, such as gestures or hand interactions. Based on the detected interactions, the MEC server generates corresponding haptic feedback, triggering precise force and vibration responses on the recipient’s gloves. During the same time, the holographic video content is processed and sent to the XR helm of the recipient via a 5G downlink. To ensure a seamless and coherent experience, timestamp-based synchronisation is applied to all data streams, maintaining temporal consistency between visual and haptic signals. Studies indicate that the temporal discrepancy between these signals must remain below 70 milliseconds to achieve a natural perceptual experience [6]. Despite the reliability of the TCP protocol, challenges such as potential data reordering and delays in wireless networks require robust synchronisation mechanisms on the MEC server. Finally, based on the source IP address, the server excludes the sender’s holographic data from the transmitted streams, ensuring that recipients receive only relevant holographic images and feedback. This integrated architecture leverages low-latency MEC processing to deliver highly interactive, immersive, and real-time holographic-haptic experiences in shared virtual spaces. In this architecture, synchronisation issues arise in two forms: desynchronisation among multiple sensory data streams from the same user during transmission and desynchronisation between frames of multiple users. For the former, accurate alignment is essential to preserve the semantic intent of user actions. As noted in [6], the temporal discrepancy between visual and haptic signals must remain below 70 milliseconds to ensure that most users can perceive a coherent experience, such as the sensation of touching the hand of another user. For the latter, the synchronisation requirements depend on the application’s tolerance for latency, though minimising discrepancies - ideally to millisecond levels - significantly enhances immersion. These challenges stem from factors such as frame size variations, caused by changes in a user’s distance or posture relative to the camera, and inherent wireless network fluctuations. To address the potential arrival time difference between multisensory data, a fixed multisensory data composition and transmission approach ensures that nonvisual data are bound to and sent immediately after visual frames. For arrival time difference between multiple users, a QL algorithm on the mobile edge server dynamically manages synchronisation, ensuring smooth and immersive interactions. B. Multi-user frame synchronisation using QL We describe the multi-user synchronisation problem in the proposed framework and design a QL algorithm to improve synchronisation performance. Due to network uncertainty and varying frame sizes, real-time transmission may result in perceived asynchronous among users, affecting their perception and feedback actions. To address this, the edge server caches early-arriving frame from one user, and uses a pre-trained QL model to decide whether to wait for a certain period or immediately forward it and discard later arrival frame from another user, balancing transmission time and frame delivered count. This decision-making focuses on the uplink transmission time, where bandwidth is limited due to network asymmetry. In the QL algorithm, let Ttrepresent the arrival time of the combined frames of two users at the same index at a given time t. Let Tt,i denote the i-th arriving frame of the user belonging to the frame Tt, then the state is defined as a weighted sum of the arrival times of the previous and current frames as: St=α·St−1+ (1 −α)·Tt,0(1) , where αis a weight factor (0≤α≤1). This timesequence weighting approach helps prevent scenarios where, if the previous frame experienced significant delay, relying solely on the arrival time of the frame from the firstarriving user for the current index might cause the QL algorithm to develop a strategy that leads to excessive waiting times. Also such inappropriate decisions could lead to a cascading effect, continuously extending the frame arrival time. For QL actions, we design discrete waiting time values at each step Atto be from 0ms to 120 ms, such as {0ms,10 ms,...,120 ms}, and 0 ms denotes the action not to wait. For obtaining the next state, the arrival time of the current frame is calculated by adding the waiting time to the arrival time of the first frame in the current state. This result is then weighted and combined with the arrival time of the first user’s frame at the next time step to determine the state. The state at the next time step St+1 can be expressed as: St+1 =α·(St+At) + (1 −α)·Tt+1,0(2) , where the default value of αis 0.5. The reward function is defined as the weighted sum of two components, controlled by the parameter β∈[0,1]: Rtotal(t) = β·Rwait(t) + (1 −β)·Rsync(t)(3) , where βadjusts the balance between minimising waiting time and maximising the number of frames that have successfully arrived. The waiting time reward, Rwait(t), decreases as waiting time increases and is given by Rwait(t) = 2 ·exp(γ·(St+At)) −1 exp(γ·(St+Amax)) −1(4) , where γis the growth rate parameter (default 0.05), and exp is the natural exponential function. The synchronisation award, denoted as Rsync, decreases proportionally based on the number of frames that do not arrive and is defined as: Rsync(t) = 1 N n X i=1 1(Tt,i < St+At),(5) where 1(Tt,i < St+At) = (1,if Tt,i < St+At, 0,otherwise.(6) and Nis the total number of streams. This reward design evaluates the trade-off between synchronisation success and waiting time penalties, relying observation of the current and previous time slot but allowing long-term optimisation of synchronisation and frame rate in a holographic streaming flow. It is worth mentioning that we opt for tabular-based QL over a neural network to approximate the Q-function. This choice stems from our experience in categorising states and corresponding strategies into discrete and finite categories, which keeps the Q-table size manageable. Consequently, this approach avoids the increased complexity that comes with applying deep neural network to approximate the Q value function. IV. EVALUATION RESULTS In this section, we present the evaluation results of the use cases based on the implemented framework. A. Implementation and evaluation tested User-side settings: In terms of capturing live real-time holographic data, we deploy two Microsoft Kinect DK cameras at each room, working in joint mode to obtain a multiview hologram and automatically combine them in a 6-DoF manner. To display such streaming to peer users, Microsoft HoloLens 2 [17] is adopted with a Unity-based programme to allow controllable frame transmission and rendering. To track the user’s real-time position of the body part, we deploy two steam VR base stations in each room to build the coordinate system. Each user wears an HTC VIVE 3.0 tracker for real-time position tracking. For haptic data, one SenseGlove Nova is provided per user, and we enable both force and vibration as haptic feedback. A dedicated C++ application was developed and deployed locally to interface with the hardware stack, aggregating data streams from all connected devices on the user side. The default data collection interval for the data collection proxy is 33 ms, which ensures that all types of data frames have the same quantity. Network entity implementation and deployment: The user side client is psychically connected to a 5G Customer Premise Equipment (CPE) to access the 5G network and initiates a TCP connection to the MEC server. The implemented 5G core network adopts a standalone mode and the radio access is 5G Enhanced Mobile Broadband (eMBB). An MEC server is deployed at the network edge with dual Nvidia A6000 video card and other matching hardware configurations. The operating system is Windows 11. The MEC server hosts an enhanced Livescan3d codebase [15] as the content server, with other functions including haptic information processing, delay monitoring, content filtering and calibration can be done in the same application. QL learning rate is set to 0.01, and the discount factor is 0.9. An epsilon-greedy exploration strategy is used. The initial state is set to 100 ms. The number of episodes and steps per episode is set to 200. The samples are pre-collected and stored offline, originating from real-world 5G networks and devices. Performance measurement: In interactive multi-user scenarios, our tests primarily focus on application latency. On each user’s local client, the arrival timestamp of different sensory data can be logged, allowing for the calculation of the time differences between visual-haptic data. Additionally, the behaviours of two users that occur simultaneously can be assessed by examining the sequence numbers and timestamps of the received frames, thus identifying differences between users. Moreover, the MEC server can monitor the real-time throughput of each user, measured in frames per second (FPS), based on the interaction of frame requests and acknowledgements over reliable connections. Fig. 2: Measured application performance over different resolution levels Fig. 3: Visual-Haptic frame delay difference under different user behaviour B. Evaluation results of use case interactive social event We first present the overall performance of this application, including FPS and throughput (see Fig. 2). Although our two local users are in different rooms, they share the same 5G access base station. As a result, only at low resolutions, such as Normal-Field-of-View (NFOV) , can the FPS exceed 20. When the resolution increases, only High Definition (HD) can achieve FPS above 10. The difference in FPS can be explained by their required bandwidth demands. Except for NFOV, all other levels of clarity occupy all the bandwidth when two users access the 5G uplink network simultaneously, resulting in a decrease in FPS. Especially, while higher resolution levels like Full High Definition (FHD) and Wide Quad High Definition (WQHD) provide superior visual clarity, their performance degrades to a point where it adversely affects the perceptible user experience, due to the increased computational load and bandwidth consumption that higher resolutions entail. In addition, we observe the time difference between visual and haptic frames for a single user. An intuitive observation is that if we design specific protocol behaviours at both the client side and the edge server side to ensure the sequential transmission of the two types of frames, the time difference between different sensory frames can be kept under 20 ms at any resolution. This time difference arises from the inherent uncertainty of the wireless uplink and downlink transmission over the 5G link. However, a slight delay in the delivery of haptic frames allows enough processing time so that visual frames can be displayed on the headset, typically in about 40 milliseconds [14]. Since it takes roughly 20 milliseconds for haptic frames to travel from the client to the gloves, under these latency conditions, both types of frames can be perceived by the user within a 50-millisecond window. This ensures that sensory feedback remains perceptible and synchronised. Considering two users moving and performing actions at different speeds within a given space, there are several uncertainties in real-time user motion capture (see Fig. 3). First, the size of the captured holographic image frames can vary due to changes in the user’s distance from the camera or being obstructed by objects. Additionally, the performance of wireless access provided by 5G base stations can also fluctuate due to the angle of the user’s receiving end and obstructions. However, the bandwidth required for the NFOV level remains relatively low, allowing for a higher FPS even when the upstream bandwidth is reduced to approximately 100Mbps, while also ensuring that the delay difference between multisensory frames remains unchanged. In contrast, the HD resolution level is more significantly affected. In particular, when users move rapidly, the collection and transmission delays of visual frames lead to a desynchronisation between the timestamp indices of visual and haptic frames, resulting in an increased gap between the two. C. Evaluation of synchronisation between users Next, we describe the time difference between two users after applying synchronisation, focussing on a scenario where the resolution is NFOV and the users remain stationary. We first compare frame arrival times during interactions between two users after applying QL synchronisation, as well as the gaps between frames from each user with the same frame timstamp index (see Fig. 4). We observed that, under strategy-driven waiting times, most frames were constrained within 100 ms, whereas the gap between users was limited to within 80 ms. In other words, a minimum of 10 FPS could be guaranteed, and in this case there was no significant delay in the reception of the same scene by both users. Furthermore, we observed that among the 200 frames sampled per user, approximately 10% to 15% of the frames were dropped. However, there were no cases of consecutive frame drops, which to some extent ensured the semantic continuity conveyed by several consecutive frames. This was attributed to our QL design: If a shorter or no-wait was executed for the current frame, the value of next state became smaller, making it more likely to select a longer wait time to capture late-arriving frames, thereby avoiding consecutive frame drops. Fig. 4: Synchronization performance between two users Fig. 5: Impact of different reward weight settings For different reward weights (the βin equation 3), adopted in various states (see Fig. 5), the overall trend shows a gradual reduction in waiting time as the state increases. However, for the states of 10 and 120 ms, we observe that they do not conform to this pattern. This is because, in cases where the current state value is relatively small, the learnt strategy tends to leverage the frame that arrives quickly to maintain smooth network transmission. This avoids the risk of the early-arriving frame being delayed excessively while waiting for the frame from the other user. Similarly, when both sides’ frames are already delayed, the strategy instead opts for a slight wait. This is because discarding the frame from the other user that has not yet arrived will not produce a significantly better overall frame arrival rate. In such cases, a brief wait can significantly improve the synchronisation performance between users. Next, we compare Fig. 6: Impact of reward weight on actions the adjustment of the reward weight parameters to balance the frame delay and frame synchronisation performance. As clearly shown in Fig. 6, once the frame delay component in the reward function is emphasised, the model automatically adopts shorter waiting durations in various states, leading to a rapid reduction in the average frame delay while increasing the number of dropped frames. Conversely, when the reward weight is linearly adjusted to prioritise frame synchronisation, corresponding changes occur in frame arrival times and dropped frames in a linear manner. This conclusion enables network administrators to effectively adjust the focus of the system based on the actual application scenario, whether to prioritise real-time dynamic holographic frame rates or to emphasise synchronised interactive perception of the same scene for both parties. V. CONCLUSION This paper tackles the challenge of supporting visual and tactile data for XR applications, vital for user experience in social and industrial settings. We propose a 5G edge computing architecture to reduce latency, manage visualhaptic data, and enable multi-user remote interaction. A reinforcement learning algorithm synchronises user frames for real-time consistency. A handshake scenario in a holographic environment demonstrates the architecture’s ability to deliver immersive XR experiences. VI. ACKNOWLEDGEMENT This project is sponsored by the EU Horizon Europe SPIRIT project (Grant Agreement 101070672, https://www.spirit-project.eu/). The authors also acknowledge the support of 6GIC industry members (6GIC) (http://www.surrey.ac.uk/5gic) for this work. REFERENCES [1] Lema, Maria A., et al. “5G case study of Internet of Skills: Slicing the human senses.” 2017 European Conference on Networks and Communications (EuCNC). IEEE, 2017. [2] Zhang, Anlan, et al. “YuZu:Neural-Enhanced Volumetric Video Streaming.” 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22). 2022. [3] Yang, Mengyu, et al. “A Comparative Measurement Study of Point Cloud-Based Volumetric Video Codecs.” IEEE Transactions on Broadcasting (2023). [4] Babaei, Aptin, Parham M. Kebria, and Saeid Nahavandi. “5G for Lowlatency Human-Robot Collaborations; Challenges and Solutions.” 2022 15th International Conference on Human System Interaction (HSI). IEEE, 2022. [5] Polachan, Kurian, et al. “Assessing Quality of Control in Tactile Cyber–Physical Systems.” IEEE Transactions on Network and Service Management 19.4 (2022): 5348-5365. [6] Di Luca, Massimiliano, and Arash Mahnan. “Perceptual limits of visual-haptic simultaneity in virtual reality interactions.” 2019 IEEE World Haptics Conference (WHC). IEEE, 2019. [7] Polachan, Kurian, et al. “TCPSbed: A Modular Testbed for Tactile Internet-Based Cyber-Physical Systems.” IEEE/ACM Transactions on Networking 30.2 (2021): 796-811. [8] Watkins, Christopher JCH, and Peter Dayan. “Q-learning.” Machine learning 8 (1992): 279-292. [9] SenseGlove Nova 1.X https://senseglove.gitlab.io/SenseGloveDocs/novaglove.html?highlight=force%20feedback [10] Caserman, Polona, et al. “Real-time body tracking in virtual reality using a VIVE tracker.” Virtual Reality 23 (2019): 155-168. [11] Han, Bo, et al. “CoMIC: A collaborative mobile immersive computing infrastructure for conducting multi-user XR research.” IEEE Network (2022). [12] Weber, Mitchell, et al. “Dynamic Pose Tracking Accuracy Improvement via Fusing HTC VIVE Trackers and Inertia Measurement Units.” International Journal of Precision Engineering and Manufacturing 24.9 (2023): 1661-1674. [13] Lwowski, Jonathan, et al. “HTC VIVE tracker: Accuracy for indoor localization.” IEEE Systems, Man, and Cybernetics Magazine 6.4 (2020): 15-22. [14] Long, Zijian, Haiwei Dong, and Abdulmotaleb El Saddik. “Interacting with New York city data by HoloLens through remote rendering.” IEEE Consumer Electronics Magazine 11.5 (2022): 64-72. [15] Qian, Peng, et al. “Remote Production for Live Holographic Teleportation Applications in 5G Networks.” IEEE transactions on broadcasting 68.2 (2022): 451-463. [16] Zhu, Yuanwei, et al. “A semantic-aware transmission with adaptive control scheme for volumetric video service.” IEEE Transactions on Multimedia (2022). [17] Microsoft HoloLens 2 [Online]: https://www.microsoft.com/engb/hololens/