PaperlayerView exact revision on arXiv

Source-linked AI summary

Edge Computing and its Application in Robotics: A Survey

Nazish Tahir, Ramviyas Parasuraman

arXiv:2507.00523v1cs.ROcs.DCcs.NI

TL;DR

Edge robotics faces gaps in context-aware offloading, handover, and interoperability. This review surveys current work and concludes that the field remains in its early stages, with persistent practical-implementation challenges.

  • Problem

    Edge robotics faces limited attention to context-aware offloading, handover risks, and interoperability concerns in heterogeneous systems.

    Source: B. Context Awareness · p. 14, C. Handover Mechanism · p. 14, F. Heterogeneity · p. 14

  • Method

    The paper reviews the current state of edge robotics and its use of edge platforms for robotic applications.

    Source: §VII · p. 15

  • Results

    The review finds that edge robotics remains in its early stages, although studies use edge platforms to address robotic computational demands.

    Source: §VII · p. 15

  • Takeaways & Limitations

    Further work can address persistent practical-implementation challenges in edge robotics.

    Source: §VII · p. 15

    Evaluations often fall short of real-world operational complexity, relying on idealized networks, deterministic robot behavior, or fixed layouts.

    Source: §V · p. 13

Abstract

from arXiv · show

The Edge computing paradigm has gained prominence in both academic and industry circles in recent years. By implementing edge computing facilities and services in robotics, it becomes a key enabler in the deployment of artificial intelligence applications to robots. Time-sensitive robotics applications benefit from the reduced latency, mobility, and location awareness provided by the edge computing paradigm, which enables real-time data processing and intelligence at the network's edge. While the advantages of integrating edge computing into robotics are numerous, there has been no recent survey that comprehensively examines these benefits. This paper aims to bridge that gap by highlighting important work in the domain of edge robotics, examining recent advancements, and offering deeper insight into the challenges and motivations behind both current and emerging solutions. In particular, this article provides a comprehensive evaluation of recent developments in edge robotics, with an emphasis on fundamental applications, providing in-depth analysis of the key motivations, challenges, and future directions in this rapidly evolving domain. It also explores the importance of edge computing in real-world robotics scenarios where rapid response times are critical. Finally, the paper outlines various open research challenges in the field of edge robotics.

I. INTRODUCTION

Edge computing places computation, storage, and network services near data sources, supporting faster processing and low-latency services. In robotics, edge resources address onboard limits and the delays of remote cloud computing for time-sensitive tasks.

I. INTRODUCTION

  • Edge computing moves computation closer to data sources and provides computing, storage, and network services near users.The paradigm aims to support low-latency, high-bandwidth services and real-time processing.Source: §I · p. 1
  • Distributed, parallel edge processing can reduce latency, bandwidth demands, and energy use compared with centralized cloud processing.The paper also describes faster data processing and transmission as benefits of proximity to data sources.Source: §I · p. 1

A. Applications of Edge Computing in Robotics

  • Robots may lack onboard capacity for large-scale real-time data processing, motivating offloading of resource-intensive tasks.Remote resources can provide global storage and processing, but mobile robots need efficient solutions with minimal latency.Source: A. Applications of Edge Computing in… · p. 2
  • Cloud robotics offers scalable processing, but remote, multi-hop cloud connections can add latency that is problematic for time-sensitive robotic applications.Edge infrastructure places cloud-like computing and storage at the network edge for applications such as image processing and path planning.Source: A. Applications of Edge Computing in… · p. 2

B. Survey Scope

The survey addresses the limited breadth of prior work by classifying edge robotics applications, reviewing the field, and discussing research challenges. Its literature review covers publications identified across three databases from 2015 to 2025.

B. Survey Scope

  • Prior surveys include experimental prototypes and system architectures, but one cited review lacks a broader literature review and detailed classification of edge-robotics solutions.Other cited work takes a societal and cross-domain perspective rather than focusing specifically on edge computing or multi-robot architectures.Source: B. Survey Scope · p. 2
  • The survey identifies agriculture, Industry 5.0, and healthcare service robotics as areas with prior field-specific reviews,,.Those works form part of the existing literature the survey positions itself alongside.Source: B. Survey Scope · p. 2
  • The survey combines a structured classification by application domain, comparative analysis, and discussion of open research challenges in edge robotics.It also situates edge robotics among distributed computing paradigms and reviews applications across robotic tasks.Source: B. Survey Scope · p. 3
  • The literature review searches IEEE Xplore, ScienceDirect, and the ACM Digital Library for relevant publications spanning 2015 to 2025.Search terms include edge computing in robotics, multi-robot edge systems, multi-edge servers in multi-robot systems, and edge robotics.Source: B. Survey Scope · p. 3

C. Article Organization

The article moves from computing paradigms and edge robotics properties to application evaluations and open challenges. It also outlines cloud and fog robotics architectures, their uses, and research issues.

C. Article Organization

  • The article organization covers foundational paradigms, edge-robotics properties and adoption, state-of-the-art evaluations, open challenges, and conclusion.The state-of-the-art section compares benefits and drawbacks of existing frameworks.Source: C. Article Organization · p. 3

A. Cloud Robotics

  • Cloud robotics uses cloud infrastructure for robot data or code execution, enabling remote control, shared resources, and collaboration among geographically dispersed robots.Cloud systems can provide datasets, global maps, object models, machine-learning resources, and shared code repositories.Source: A. Cloud Robotics · p. 3
  • Remote cloud servers can impose connection and data-transfer latency, a drawback for time-sensitive cloud-robotics applications [14].Cloud robotics spans tasks including localization, navigation, manipulation, perception, service robotics, and human-robot interaction.Source: A. Cloud Robotics · p. 3

B. Fog Robotics

  • Fog robotics places decentralized storage, network functions, and control near robots, with a fog server and cloud system in the architecture.Multi-robot systems can request fog-layer data and use available resources without querying the cloud,.Source: B. Fog Robotics · pp. 3–4
  • Fog robotics supports offloading, monitoring, control, collaboration, mobility, and navigation; a local model cache can serve robots with restricted onboard resources.The survey also identifies data maintenance, handover delay, security, and robot battery status as fog-robotics issues.Source: B. Fog Robotics · p. 4

C. Edge Robotics

Edge robotics distributes computation across robots and network-edge resources, while related paradigms differ in where and how they process, store, and offload data. The section distinguishes edge robotics from fog, mist, MEC, and cloudlets, and notes trade-offs among locality, latency, and infrastructure capabilities.

D. MEC and similar concepts

  • Mist computing uses embedded peripheral devices to preprocess or filter sensor input before sending it to fog or cloud.These devices may use microcontrollers or FPGAs, but have restricted processing capabilities.Source: D. MEC and similar concepts · p. 4
  • Multi-Access Edge Computing (MEC) places processing and storage in the radio access network to reduce latency and increase context awareness,.MEC resources can reduce mobile-device computation and energy use by providing one-hop services near base stations.Source: D. MEC and similar concepts · p. 4
  • Cloudlets are nearby small clusters or data centers intended to improve interactive mobile-app performance under high latency or jitter requirements.A cloudlet-based emergency-management service distributes resources for multirobot systems in a smart factory.Source: D. MEC and similar concepts · pp. 4–5

E. Difference between Cloud, Edge, Fog, MEC, Cloudlets

  • Cloud computing centralizes processing, while edge, fog, mist, and cloudlets use distributed processing and may require smaller specialized equipment.Cloudlets mainly use virtual machines, whereas MEC and fog can also use lighter-weight containers.Source: E. Difference between Cloud, Edge,… · p. 5
  • Edge nodes can process robot data locally without cloud offloading, whereas fog nodes preprocess data before forwarding it to the cloud.Fog adds an intermediary layer; edge computing can avoid that additional infrastructure, while offering less scalability and resource pooling than fog.Source: C. Edge Robotics · p. 4, E. Difference between Cloud, Edge,… · p. 5

III. KEY CHARACTERISTICS OF EDGE ROBOTICS

Edge robotics draws on nearby computing resources to support time-sensitive robot applications. Its characteristics include proximity, low latency, geographical distribution, and mobility support.

A. Proximity

  • Nearby edge resources let robots use context-aware network information for offloading decisions and access remote services as a “remote brain.”The section presents these resources as useful for time-sensitive applications.Source: A. Proximity · p. 5

B. Low latency

  • Low-latency access lets robots run resource-hungry, delay-sensitive algorithms such as SLAM, navigation, object detection, and gesture recognition on edge devices.Examples of edge devices include routers, access points, base stations, and dedicated servers.Source: B. Low latency · p. 5

C. Geographical distribution

  • Edge infrastructure’s dense geographical distribution supports mobile-robot offloading and can improve resilience to network loss through faster backups and recovery.It offers cloud-like service provisioning closer to local robots.Source: C. Geographical distribution · p. 5

D. Mobility support

  • LISP supports robot mobility by separating host identity from location, allowing mobile robots to access services near their current location.The passage describes this decoupling as the basis for mobility support.Source: D. Mobility support · p. 5

E. Heterogeneity

The survey describes computational offloading approaches that distribute robotic workloads across heterogeneous edge resources and communication networks. Reported work spans SLAM, fleet control, adaptive offloading, and network-aware robotic navigation.

E. Heterogeneity

  • Edge heterogeneity spans device, server, and network platforms, including differences in hardware, software, APIs, policies, and communication technologies.These variations influence how edge services are delivered to mobile robots.Source: E. Heterogeneity · p. 6

A. Computational Offloading

  • Dey et al. dynamically offload computationally expensive SLAM components based on compute load, communication load, and energy use, reducing mean execution time versus static offloading in VM experiments.The survey notes that the strategy was not tested in real-world conditions, where robot-to-gateway transmission could bottleneck.Source: A. Computational Offloading · p. 6
  • NETMODE manages heterogeneous computing and network resources for mobile-robot manufacturing scenarios, and its edge-assisted SLAM tests report faster compute-intensive workloads with edge infrastructure.The experiments compare map-update intervals and CPU use for gmapping on local and edge resources.Source: A. Computational Offloading · p. 6
  • Li et al. use heterogeneous GPU, NPU, TPU, APU, and FPGA edge hardware to filter robot data and support real-time AI computation across large robot fleets.Their live-robot control approach processes sensor inputs and sends operation instructions while receiving cloud updates and training.Source: A. Computational Offloading · p. 6
  • Wang et al. use a UAV cluster and base station for collaborative task offloading, optimizing latency and energy in simulation; real-world performance under unpredictable conditions remains untested.An improved genetic algorithm solves the resource-optimization problem.Source: A. Computational Offloading · pp. 6–7
  • Multi-robot offloading work addresses utility-aware task decisions and decentralized scheduling, while a cloud-edge, multi-RAT testbed supports real-time offloading of vision tasks,,.The testbed combines Kubernetes, Istio, and 5G/mmWave networks, broadening the design space to include network heterogeneity.Source: A. Computational Offloading · p. 7

B. Context Awareness

The section surveys context-aware edge robotics and localization systems, including semantic navigation, network-responsive control, and distributed mapping. These approaches report benefits alongside limits such as longer navigation time or evidence from restricted test settings.

B. Context Awareness

  • Klaas et al. add semantic environmental information to edge-enabled path planning, but the cited work does not assess operational-time reductions.Related semantic-navigation approaches also use robot policies based on classified surroundings,.Source: B. Context Awareness · p. 7
  • Offloading an entire navigation stack to the edge reduced energy use and onboard computing requirements, but increased operation time.The passage identifies this as a trade-off in the microservice-based architecture.Source: B. Context Awareness · p. 7
  • Wi-Fi signal context was used to adjust robot speed, with experiments indicating smoother driving at high speeds than without that information.The approach connects communication conditions to remote robot control.Source: B. Context Awareness · p. 7
  • COTORRA emulates unpredictable network conditions for edge-based robotic testing and reported autonomous-navigation task latency below 15 ms.The testbed supports serverless plugins and orchestration involving network infrastructure and mobile robots.Source: B. Context Awareness · p. 7
  • A 6G semantic-communication framework uses a knowledge graph of object-action sequences to guide robots and transmits compressed features for real-time recognition and path identification.Synthetic and real-world datasets showed latency reductions without compromising inference accuracy.Source: B. Context Awareness · p. 7

C. Localization

  • ColaSLAM assigns map submaps to edge clusters using overlap and bandwidth conditions, reducing processing latency by up to 40% versus direct cloud offloading.It improves by up to 52% on average versus random assignment and up to 15% versus the greedy algorithm.Source: C. Localization · p. 8
  • RecSLAM distributes multi-robot laser SLAM across nearby edge servers, while a separate edge-fog architecture uses handover and prior-map exchange to reduce delay and robot energy use.The latter was tested on a small automobile prototype, but the review says the results were insufficient to demonstrate efficacy.Source: C. Localization · p. 8
  • An edge-split ORB-SLAM2 architecture keeps tracking on the robot and moves local mapping and loop closure to the edge, reducing computation and memory costs with negligible mapping-accuracy loss.Prototype experiments also report that reduced local-map update latency lowers overall VSLAM execution time.Source: C. Localization · pp. 8–9

D. Navigation

The navigation literature uses edge resources for robot path planning, collaboration, and visual navigation, with reported gains in execution time and resource use alongside limits in evaluation. Results also show that offloading strategy and image transmission can materially affect processing time.

D. Navigation

  • An edge-assisted manufacturing robot used path calculation to navigate around obstacles, while another proposed edge navigation module lacked benchmark comparisons for its direction-tracking and operation-time tests.The missing comparisons limited the conclusions that could be drawn from the latter study.Source: D. Navigation · p. 9
  • Edge computing reduced execution time, upload time, CPU use, and memory use in experiments comparing local, cloud, and edge navigation for heterogeneous robots.The study involved a mobile robot and manipulator collaborating on navigation and object identification for grasping.Source: D. Navigation · p. 9
  • Li et al. report better visual-navigation performance in non-interference and complex-noise scenes, but provide no edge-platform data to establish the edge device’s contribution.The approach combines deep image understanding, a cascaded convolutional network, hybrid dilated convolution fusion, and a modified Hough transform.Source: D. Navigation · p. 9
  • Image compression reduced image size and network bandwidth by an order of magnitude without compromising odometry accuracy, and the study reports benefits for edge offloading.The tests examined network latency and odometry accuracy using the EuRoC dataset.Source: D. Navigation · pp. 9–10
  • Partial image offloading outperformed complete offloading in processing time across resolutions, with the gap widening at higher resolutions.When server power was twice the drone’s, partial offloading was faster than onboard processing; full offloading was not.Source: D. Navigation · pp. 10–11
  • Higher uplink speeds can meet full-offloading image-transfer requirements, but sending extracted features uses less transmission rate than sending entire images.Partial offloading still requires some onboard processing.Source: D. Navigation · p. 11
  • EARN adaptively switches between local and edge motion planning under communication and computation constraints, but depends on low-latency connectivity and complex optimization.The framework reports strong performance in simulated and real-world settings, with potential deployment challenges in dynamic or resource-constrained environments.Source: D. Navigation · p. 11

E. Minimizing task latency

Latency-focused work places SLAM and task computation at or across edge servers, while accounting for robot mobility and resource constraints. Reported approaches reduce processing latency or enable task switching, but handovers and connectivity remain relevant deployment concerns.

E. Minimizing task latency

  • The proposed edge architecture computed SLAM with less processing latency than two benchmark techniques in a collaborative robot-and-manipulator automation cell.An edge-generated map supported a mobile robot delivering a package using pose estimates from the edge device.Source: E. Minimizing task latency · p. 11
  • A master robot acting as an edge device makes decisions from slave-robot sensing, while resource-management schemes balance or reduce slave energy use and support on-time tasks.The second scheme uses slave residual energy and may consume more energy than the first, but is described as more durable in unpredictable wireless conditions.Source: E. Minimizing task latency · p. 11
  • Robot mobility can trigger access-point handovers and service migration, increasing latency and impairing tasks such as autonomous navigation.A simulated DDPG framework jointly allocates communication and computation across access points to support task switching among edge servers.Source: E. Minimizing task latency · p. 11
  • FogROS2-LS dynamically selects cloud or edge services to meet latency requirements, offloading state estimators and feedback controllers without modifying existing ROS2 applications.It switches among identical services to select an optimal deployment for the latency requirement.Source: E. Minimizing task latency · p. 11

F. Resource Optimization

Resource-optimization work spans parallel lidar processing, adaptive localization offloading, and task scheduling. The approaches weigh processing and resource use against navigation performance, mission duration, and energy needs.

F. Resource Optimization

  • The optimized FPGA lidar-odometry design outperformed its initial unoptimized design in resource usage and concurrency.The FPGA use targets real-time scan processing with minimal computational resources and parallelized data processing.Source: F. Resource Optimization · p. 12
  • A utility-based offloading scheme combining local odometry and edge-based vision localization outperformed two alternatives while providing convergence guarantees and keeping mission duration low.The alternatives were exclusively local and exclusively remote implementations.Source: F. Resource Optimization · p. 12
  • A collaborative vehicular-edge scheme probabilistically assigns tasks and uses contract-based incentives for parking vehicles to maximize subjective utility under information asymmetry.The authors report numerical experimental evidence of the scheme’s efficacy and efficiency.Source: F. Resource Optimization · p. 12
  • A DRL scheduling strategy jointly optimizes task execution, robot trajectory, and energy use, with evaluations reporting reduced delays and energy use against baselines.The approach uses an MDP for adaptive choices including local processing, offloading, and device-to-device task sharing; it remains dependent on simulation scenarios.Source: F. Resource Optimization · p. 12

G. Minimizing energy consumption

Energy-focused work includes robot task offloading, contextual resource allocation, and collaborative edge training. The survey also identifies limits in experimental realism, coverage, and deployment readiness that constrain evaluation of edge robotics.

G. Minimizing energy consumption

  • Wang et al. use deep reinforcement learning to schedule mobility-aware offloading for a robot swarm, reporting lower energy consumption and delay in simulation.The system uses one edge server, leaving the effects of deploying multiple edge devices unassessed.Source: G. Minimizing energy consumption · p. 12
  • Farahbaksh et al. use a control loop to capture contextual information for offloading decisions and evaluate energy, execution time, resource use, network use, latency, module size, and offloading intervals.Source: G. Minimizing energy consumption · p. 12
  • Zeng et al. propose collaborative edge training that pools idle trusted edge-device resources to address long-range backhaul strain from centralized training.Their work analyzes energy-oriented scheduling and evaluates parallelism methods using realistic testbeds.Source: G. Minimizing energy consumption · p. 12

V. RESEARCH GAPS AND LIMITATIONS

  • Dey et al. used virtual machines without physical robots and assumed ideal network conditions; the NETMODE testbed used resource-limited Raspberry Pi 3 robots and omitted several key performance measures.NETMODE also lacks dynamic task orchestration, multi-robot scalability, and resource-aware edge decisions.Source: §V · p. 13
  • Some studies use static layouts, predictable tasks, idealized communications, or controlled workloads, limiting their relevance to variable operating conditions.The survey cites proximity-based allocation that does not scale to complex environments, specialized testbed infrastructure with narrow metrics, and energy savings measured without edge-energy costs or network overheads.Source: §V · p. 13
  • COTORRA uses synthetic network variability and constrained indoor tests, while ColaSLAM assumes uninterrupted connectivity and may face scheduling delays and map-fusion reliability issues.COTORRA also uses pre-programmed robot paths and a 15 ms orchestration target; its DLT federation delay is 19 seconds.Source: §V · p. 13
  • Edge-SLAM was evaluated on two mobile devices and one edge server using replayed datasets and private Wi-Fi, limiting generalizability to more complex deployments.The survey notes that latency and architectural decomposition could exacerbate drift in longer or more complex settings, and inherited tracking sensitivity remains unaddressed.Source: §V · p. 13
  • The survey identifies idealized networks and robot behavior, small testbeds, synthetic data, and limited metrics as barriers to assessing scalability, fault tolerance, and adaptability.It calls for robust, scalable, generalizable evaluation frameworks grounded in real-world constraints.Source: §V · pp. 13–14

VI. CHALLENGES FOR FUTURE RESEARCH

The review identifies unresolved implementation challenges for edge-based robotic offloading, including security and limited context awareness. It calls for these concerns to inform offloading decisions.

VI. CHALLENGES FOR FUTURE RESEARCH

  • Offloading can use remote resources for latency-sensitive, compute-intensive robotic operations, but network conditions introduce delay and energy challenges.The review identifies these as open areas for further investigation.Source: §VI · p. 14

A. Security

  • Security risks include data modification, denial of service, and information leakage, yet prior studies did not make security a primary optimization objective.The review recommends incorporating security into offloading decisions.Source: A. Security · p. 14

B. Context Awareness

  • Context awareness has received limited attention, although offloading decisions should account for server availability, network quality, and device workload.The review says context should inform every offloading stage, including program partitioning.Source: B. Context Awareness · p. 14

C. Handover Mechanism

The review describes risks to robotic task execution during handovers, network failures, and offloading, alongside interoperability concerns in heterogeneous systems. It concludes that edge robotics remains early-stage and that this broad review is not exhaustive.

C. Handover Mechanism

  • During edge-device handover, data loss can leave robots idle or halted, adding latency, cost, and potential safety hazards during navigation.The review calls for efficient handover mechanisms to avoid these outcomes and task failures.Source: C. Handover Mechanism · p. 14

D. Network failure

  • Connectivity loss can leave robots waiting for communication to resume; local execution may be problematic when algorithms are split across resources.The passage identifies complete network failure and temporary connectivity loss as possible causes.Source: D. Network failure · p. 14

E. Fault tolerance

  • For critical scenarios such as search and rescue, the review calls for local execution or backup remote resources to avoid delays when offloading fails.Failures may stem from weak robot connections or remote edge-server problems.Source: E. Fault tolerance · p. 14

F. Heterogeneity

  • Integrating heterogeneous edge systems with robots can create interoperability concerns that hinder deployment in heterogeneous or multirobot systems.Offloading failures can also increase energy use or execution and transmission latency, motivating recovery strategies.Source: F. Heterogeneity · p. 14

VII. CONCLUSION

  • The review is broad but not exhaustive, particularly lacking a detailed literature classification based on Fog computing technologies.It characterizes edge robotics as early-stage and identifies resource management, scalable architectures, failure recovery, and security as persistent implementation challenges.Source: §VII · p. 15

Related papers

Loading 2507.00523v1…