PaperlayerView exact revision on arXiv

Source-linked AI summary

Semi-Autonomous Prosthesis Control Empowered by 5G and Mobile Edge Computing

Ozan Karaali, Hossam Farag, Strahinja Dosen, Cedomir Stefanovic

arXiv:2608.31021v1eess.SY

TL;DR

Wearable prostheses benefit from computer-vision-based grasp assistance, but their limited onboard resources make real-time vision processing impractical. This paper develops and evaluates a 5G-connected MEC prosthesis that offloads grasp planning to an edge server, finding that all tested 5G configurations match wired Ethernet and outperform manual control and local processing. The study’s scope is limited by laboratory testing and able-bodied participants using a prosthesis simulator.

  • Problem

    Wearable prostheses have insufficient processing resources for real-time computer vision, while existing lab-tethered systems limit evaluation beyond the laboratory.

  • Method

    The study implements and experimentally evaluates a 5G-connected semiautonomous prosthetic hand that offloads vision-based grasp planning to MEC resources.

  • Results

    All 5G configurations matched wired Ethernet across task completion time, failure rate, and subjective workload, while significantly outperforming manual control and on-device processing.

  • Takeaways & Limitations

    5G edge/cloud offloading is established as a practical path for deploying sophisticated compute-intensive prosthesis control.

  • Takeaways & Limitations

    Testing occurred in a controlled laboratory with able-bodied participants using a prosthesis simulator, so diverse environments and clinical users require further evaluation.

Abstract

from arXiv · show

Prosthetic hands equipped with cameras can use computer vision to plan grasps automatically, reducing cognitive effort. However, running modern vision models on wearable devices is impractical due to power and processing constraints. We present the first prototype of a 5G-connected mobile edge computing (MEC)-enabled semi-autonomous prosthetic hand, which streams RGB-D images to an edge server for real-time grasp planning. Thirteen able-bodied participants performed pick-and-place tasks under six conditions: manual EMG control, on-device inference, wired Ethernet connectivity, and three 5G connectivity configurations (private 20 MHz network, private 100 MHz network, and a commercial 5G link) to the server. All network-based conditions performed similarly, achieving task times around 8.6 s (34% faster than manual control), failure rates of 20-38%, and 62% lower overall workload. On-device processing performed the worst with 10.3 s task time and a 76% failure rate due to slow embedded inference (3 fps vs. 6-20 fps over the network). Network latencies remained below 180 ms for private 5G and 270 ms for commercial 5G. All 5G configurations, including bandwidth-constrained and commercially variable networks, matched wired Ethernet performance while significantly outperforming both manual control and local processing, establishing 5G edge-offloading as a practical path to deploying compute-intensive prosthesis control.

I. INTRODUCTION

Semi-autonomous prosthesis control shifts users from low-level motor execution toward high-level intent specification, but wearable devices cannot practically run the required real-time vision models. This work evaluates whether 5G connectivity and edge offloading can provide performance comparable to wired connectivity while improving on manual and local processing.

  • Motivation: Surface EMG control requires distinct muscle activation patterns for grip selection, wrist orientation, and hand aperture, creating substantial cognitive burden.The burden can become exhausting when controlling prostheses with many degrees of freedom.
  • Motivation: Semi-autonomous control uses cameras or depth sensors and computer vision to automate grasp configuration while the user specifies what to grasp and when.The system can compute grip type, wrist rotation, and aperture from visual input.
  • Challenge: Real-time computer vision exceeds the processing and data resources of compact prosthetic devices, while lab-tethered computers limit translation beyond the laboratory.Deep neural-network detection and segmentation require GPU-class hardware unavailable in the prosthesis.
  • Approach: MEC addresses this constraint by offloading intensive perception tasks to nearby servers over wireless networks, reducing dependence on wearable computation.Prior cloud offloading reduced smartphone power consumption by 40% and achieved classification delays of 90–370 ms.
  • Study: The study evaluates a first 5G-connected semiautonomous prosthetic hand with thirteen participants across manual, on-device, Ethernet, private-5G, and commercial-5G conditions.Performance was assessed using task completion time, failure rate, and subjective workload.
  • Findings: All 5G configurations matched wired Ethernet performance and outperformed manual control, whereas on-device processing performed worst because of insufficient inference speed.The reported conclusion identifies 5G-connected prosthesis control as viable and advantageous.

II. 5G-CONNECTED MEC-ENABLED PROSTHESIS PROTOTYPE

The prototype combines a multi-articulated myoelectric hand, a dorsal RGB-D camera, and a Raspberry Pi control unit with direct 5G connectivity. These components support camera-based sensing and communication with the remote processing system.

  • A. System components: The Ottobock Michelangelo hand provides powered fingers, a repositionable thumb, and palmar and lateral grasp patterns.It is a multi-articulated myoelectric prosthesis.
  • A. System components: An Intel RealSense D405 camera mounted on the hand captures synchronized 640×480 color and depth imagery at 30 fps.Its optimal operating range is 7–50 cm with an 87°×58° field of view.
  • 2) Vision System: The architecture streams camera frames from the prosthesis through the control unit and 5G network to an inference server, which returns grasp commands for hand control.EMG electrodes enable the user to trigger automatic control and orient the prosthesis manually.
  • A. System components: The Raspberry Pi 5 handles EMG acquisition, camera streaming, motor commands, and network communication through a Quectel 5G modem.The modem provides direct cellular connectivity without intermediate WiFi or router hops.

4) 5G Private Network:

The prototype uses a remote vision-and-grasp pipeline connected through a 5G edge infrastructure. It continuously processes camera frames, validates detections, computes grasp parameters, and lets the user trigger automatic preshaping before manual control.

  • 4) 5G Private Network:: The edge infrastructure combines an SDR-based 5G base station, virtualized 5G core functions, and a GPU-equipped inference server.The server hosts both network functions and the computer-vision pipeline.
  • B. System operation: The system streams frames continuously over 5G using pipelined transmission, achieving approximately 6–16 fps depending on configuration.Ethernet provides approximately 20 fps.
  • B. System operation: The server analyzes frames to estimate object size and orientation, then returns grasp type, aperture, and wrist rotation as a JSON command.Commands are sent when valid objects are detected in the workspace.
  • B. System operation: The user triggers automatic preshaping with a brief wrist-flexor activation, after which the hand rotates, selects the grasp, and opens to the computed aperture.The automatic configuration is based on the latest grasp parameters received from the edge server.
  • B. System operation: After preshaping, the prosthesis enters manual control, allowing proportional EMG reconfiguration or hand closure around the object.A subsequent co-contraction returns the system to automatic mode after object release.
  • C. Computer Vision and Grasp Selection Pipeline: YOLOv11-large performs instance segmentation, PCA estimates object geometry, and rule-based logic selects the grasp configuration.The pipeline also validates depth range and aperture limits before invoking grasping.

III. EXPERIMENTAL ASSESSMENT

The experimental assessment compares manual control, local inference, wired Ethernet, and three private or commercial 5G configurations. The design isolates how network characteristics affect prosthesis control performance.

  • A. Tested Network Configurations: Six conditions were tested to systematically assess how network characteristics affect prosthesis control performance.The conditions included manual control, on-device inference, Ethernet, private 20 MHz, private 100 MHz, and commercial 5G.
  • A. Tested Network Configurations: Manual Control served as the clinical myoelectric baseline for quantifying the benefits of a smart connected prosthesis.Participants used a 2-channel direct and proportional scheme with switching.
  • A. Tested Network Configurations: On-Device Inference ran the vision pipeline locally on a Raspberry Pi 5 using YOLOv11-nano, testing fully wearable processing without network dependence.This condition compared a simpler embedded model with the full model used for network offloading.
  • A. Tested Network Configurations: Ethernet provided a Gigabit wired direct connection to the server as a lower-latency baseline for wireless conditions.It established a reference for network-based performance.
  • A. Tested Network Configurations: Commercial 5G operated over a public non-standalone network whose traffic traversed the public internet to the server.This condition represented shared spectrum and variable network load.

B. Participants

Thirteen able-bodied participants performed standardized pick-and-place trials using visual cues that specified objects, grasp types, and wrist rotations. Each condition included 15 trials across object-orientation combinations, with counterbalanced condition order.

  • Participants: 13 able-bodied participants completed the prosthesis experiments.Participants were 10 male and 3 female, aged 24–63 years.
  • Task and Procedure: Visual cues specified the target object, required palmar or lateral grasp, and desired wrist rotation.The cues guided each pick-and-place trial.
  • Task and Procedure: Participants grasped each object, transported it to a collection box, and released it inside.The setup placed the visual-cue laptop on the left, target object centrally, and collection box on the right.
  • Task and Procedure: 15 trials per condition covered object-orientation combinations in a fixed sequence.Each trial allowed up to three grasp attempts; failed trials included unsuccessful attempts or timeouts.
  • Experimental Design: Conditions were counterbalanced across participants using a Latin Square design.This controlled the order in which participants experienced the six conditions.

D. Outcome Measures

Performance was assessed using task completion time, failure rate, subjective workload, and connectivity-related technical measures. These measures captured task efficiency, success, user burden, and network-processing behavior.

  • Primary Metrics: Task completion time Tc measured elapsed time from cue presentation to object release.It represented the duration of each complete pick-and-place trial.
  • Primary Metrics: Failure rate Fr counted trials with rotation errors, incorrect grasp selection, grasp failures, or object drops.Rotation errors were deviations greater than 30° from the target angle.
  • Primary Metrics: NASA-TLX workload averaged six 0–100 subscales: mental, physical, temporal, performance, effort, and frustration.The study used raw, unweighted scoring after each condition.
  • Technical Metrics: Connectivity monitoring included round-trip latency, effective bandwidth, inference frame rate, and server-side GPU processing time.These technical parameters were logged for network-connected conditions.

E. Statistical Analysis

The analysis used nonparametric tests to compare conditions because completion times violated normality in some conditions and the sample size was small. Network-performance results also contrasted edge-server and on-device processing behavior.

  • Statistical Analysis: Nonparametric testing was used uniformly because completion times violated normality in three of six conditions and N = 13 limited normality-test power.The analysis used Friedman omnibus tests and Wilcoxon signed-rank post-hoc tests with Bonferroni correction.
  • Network Performance: Edge-server GPU processing averaged approximately 22 ms, compared with 429 ms for Raspberry Pi on-device inference.On-device processing was roughly 20 times slower despite avoiding network latency.
  • Network Performance: On-device processing sustained approximately 3 fps and exhibited P95 tail latency of 2249 ms.CPU contention among inference, camera capture, and EMG processing caused frame dropping.

B. Task Performance

Connected prosthesis conditions improved task speed and reduced workload relative to manual control while maintaining comparable failure rates. All 5G configurations performed comparably to Ethernet, whereas on-device processing was substantially less reliable.

  • B. Task Performance: All connected conditions significantly shortened task completion time versus manual control.Mean completion time decreased from 12.99 s manually to 8.27 s with Ethernet and 8.26 s with 5G-100, a reduction of about 36%.
  • B. Task Performance: Connected-condition failure rates ranged from 20–38% and did not differ significantly from manual control.On-device inference reached 75.9% failure, significantly worse than every connected condition and manual control.
  • B. Task Performance: Ethernet, private 5G, and commercial 5G showed no significant differences in task time, failure rate, or workload.This indicates comparable performance across wired, private-wireless, and commercial-wireless connectivity.
  • C. Subjective Workload: Manual control produced the highest workload, with a mean score of 52.1 driven mainly by mental demand and effort.Mental demand averaged 67.7 and effort averaged 60.0.
  • C. Subjective Workload: Ethernet reduced mean workload to 20.0, a 62% reduction from manual control.Private 5G produced comparable workload scores of 24.2 and 24.6, while commercial 5G scored 29.3 without reaching significance versus manual control.
  • V. DISCUSSION: The study’s prototype matched ideal wired performance and exceeded conventional manual control and on-device processing.The authors conclude that MEC-assisted shared control reduced perceived cognitive workload while accelerating computation.

A. Feasibility of 5G-Connected Prosthetic Control

5G-connected prosthesis control matched wired Ethernet across outcome measures despite higher network latency, while on-device processing failed because embedded inference was too slow. The shared-control pipeline absorbed network delay through continuous streaming and pre-computed grasp solutions.

  • A. Feasibility of 5G-Connected Prosthetic Control: All 5G configurations matched wired Ethernet in task completion time, failure rate, and subjective workload.This included bandwidth-constrained private 20 MHz and variable commercial 5G networks.
  • A. Feasibility of 5G-Connected Prosthetic Control: 65–234 ms network latency did not produce measurable performance degradation because grasp solutions were pre-computed from continuously processed frames.The system processed frames at 6–20 fps and used an already-computed result when preshaping was triggered.
  • A. Feasibility of 5G-Connected Prosthetic Control: The connected prosthesis reduced workload relative to manual control, whose mental demand and effort were highest.The subjective workload figure places connected conditions near the center of the polar plot, indicating low workload.
  • A. Feasibility of 5G-Connected Prosthetic Control: 3 fps on-device inference and 429 ms processing made grasp solutions stale or unavailable when users triggered preshaping.Network-offloaded conditions processed at 6–20 fps, whereas on-device processing performed worst despite avoiding network latency.
  • A. Feasibility of 5G-Connected Prosthetic Control: 75.9% failure for on-device ODML reflected temporally unavailable detections rather than model accuracy.Users repeatedly triggered without response, causing frustration and disengagement.
  • A. Feasibility of 5G-Connected Prosthetic Control: Current embedded hardware cannot support real-time deep-learning perception at useful frame rates, keeping network offloading advantageous until embedded accelerators improve.The passage describes an order-of-magnitude inference-performance gap between embedded CPUs and discrete GPUs.

C. Comparison with Prior Work

The connected architecture achieves task performance comparable to prior tethered or embedded systems while supporting more computationally intensive processing and practical wireless deployment. Its edge-based design also provides modularity, portability, and deployment trade-offs that remain relevant outside the laboratory.

  • C. Comparison with Prior Work: 8.6 s connected-prosthesis task completion aligns with the 9–11 s reported for MYO-PACE embedded processing.The network-offloaded architecture achieves comparable performance while enabling more computationally intensive vision models.
  • C. Comparison with Prior Work: Direct 5G-to-edge connectivity avoids the smartphone intermediary used in prior cloud-offloaded EMG classification.The architecture instead performs vision-based grasp planning through a MEC-enabled base station and inference server.
  • C. Comparison with Prior Work: The vision pipeline combines YOLO detection, PCA-based measurement, and rule-based grasp selection.The authors intentionally kept this pipeline straightforward because the contribution is system architecture assessment rather than a new grasp-planning method.
  • C. Comparison with Prior Work: The framework can incorporate learned grasp planners, dexterous coordination, and multimodal sensing without structural changes.Only the inference server and transmitted sensor data need modification for these alternatives.
  • C. Comparison with Prior Work: The control unit operated portably for approximately 5–6 hours using a 20,000 mAh, 74 Wh power bank.Development tests confirmed operation while the user moved in the laboratory.
  • C. Comparison with Prior Work: Commercial 5G showed P95 latency of 890 ms versus 287–472 ms for private networks, despite functional task performance.The passage identifies congestion and routing as sources of variability and discusses network slicing and multiconnectivity as robustness mechanisms.
  • C. Comparison with Prior Work: The prototype lacks fallback to manual control during outages, so production systems should transition to EMG-only operation when connectivity is lost.Network conditions vary by location and time.
  • C. Comparison with Prior Work: RGB-D streaming captures surrounding scenes as well as target objects, creating privacy concerns for healthcare deployment.Suggested safeguards include encryption, on-device anonymization, or institutional edge computing.

E. Limitations

The study establishes technical feasibility but leaves clinical generalization, environmental robustness, and future system architecture for further evaluation. Proposed directions include clinical validation, predictive control, and hybrid on-device/edge processing.

  • E. Limitations: Testing occurred in a controlled laboratory with line-of-sight to the 5G base station, limiting direct evidence for diverse environments and mobility.Further evaluation is required under congestion, signal variation, building penetration, and handover conditions.
  • E. Limitations: Able-bodied participants using a prosthesis simulator do not fully represent prosthesis users.Longitudinal clinical studies are needed to assess everyday applicability, adoption, satisfaction, and quality of life.
  • E. Limitations: The study demonstrates technical feasibility of MEC-enabled control but does not establish clinical translation.The authors identify clinical exploration as future work.
  • E. Limitations: Hybrid architectures could run lightweight detection locally while offloading grasp planning to the edge.This design is proposed to preserve model capability while reducing latency and enabling graceful degradation during connection loss.
  • E. Limitations: Predictive control could track arm movement and gaze direction to anticipate targets and pre-compute grasp solutions.The proposed mechanism would further mask network latency.
Loading 2608.31021v1…