Source-linked AI summary
Edge Intelligence for Autonomous Driving in 6G Wireless System: Design Challenges and Solutions
Bo Yang, Xuelin Cao, Kai Xiong, Chau Yuen, Yong Liang Guan, Supeng Leng, Lijun Qian, Zhu Han
TL;DR
Autonomous vehicles face high computing demands and uncertain inference accuracy when processing large sensor streams near-real-time, motivating edge intelligence for nearby computation. The paper proposes a two-tier architecture with MTL-based offloading and segmented edge-vehicle inference, and experiments demonstrate its effectiveness while identifying limited labeled data and sensing deficiencies as open boundaries.
Problem
Autonomous vehicles must process large sensor datasets near-real-time despite high onboard computing costs, uncertain inference accuracy, limited edge resources, communication constraints, and privacy risks.
Method
A two-tier framework splits a deep neural network between vehicles and an edge server, using MTL for near-real-time offloading decisions and resource allocation plus neural-network segmentation for joint inference.
Results
The MTL model solves the MINLP in less than one-thousandth of the sBB scheme's time for 2 to 8 autonomous vehicles, while edge inference becomes advantageous when the bad-data ratio η ≥0.3.
Takeaways & Limitations
The framework combines efficient offloading, privacy preservation, and improved inference accuracy while meeting delay constraints in the demonstrated case study.
Takeaways & Limitations
The paper identifies insufficient correctly labeled training data and sensing deficiencies from video-only cameras as open research challenges.
Abstract
from arXiv · showhide
In a level-5 autonomous driving system, the autonomous driving vehicles (AVs) are expected to sense the surroundings via analyzing a large amount of data captured by a variety of onboard sensors in near-real-time. As a result, enormous computing costs will be introduced to the AVs for processing the tasks with the deployed machine learning (ML) model, while the inference accuracy may not be guaranteed. In this context, the advent of edge intelligence (EI) and sixth-generation (6G) wireless networking are expected to pave the way to more reliable and safer autonomous driving by providing multi-access edge computing (MEC) together with ML to AVs in close proximity. To realize this goal, we propose a two-tier EI-empowered autonomous driving framework. In the autonomous-vehicles tier, the autonomous vehicles are deployed with the shallow layers by splitting the trained deep neural network model. In the edge-intelligence tier, an edge server is implemented with the remaining layers (also deep layers) and an appropriately trained multi-task learning (MTL) model. In particular, obtaining the optimal offloading strategy (including the binary offloading decision and the computational resources allocation) can be formulated as a mixed-integer nonlinear programming (MINLP) problem, which is solved via MTL in near-real-time with high accuracy. On another note, an edge-vehicle joint inference is proposed through neural network segmentation to achieve efficient online inference with data privacy-preserving and less communication delay. Experiments demonstrate the effectiveness of the proposed framework, and open research topics are finally listed.
I. INTRODUCTION
The paper positions edge intelligence and 6G as a way to support reliable autonomous driving despite high onboard computing demands, imperfect inference, and constrained communication and computing resources. It proposes a two-tier framework combining intelligent offloading with privacy-preserving joint inference.
- 6G networks are expected to provide ubiquitous vehicular connectivity with low latency and high reliability, supported by artificial intelligence.
- Edge intelligence lets autonomous vehicles offload sensing data to a nearby edge server for higher-accuracy, lower-latency inference, while bandwidth, computation, privacy, and security remain challenges.
- The proposed two-tier framework places shallow neural-network layers on vehicles and the remaining layers with a multi-task learning model at the edge server.
- Multi-task learning infers binary offloading decisions and edge-server resource allocation in near-real-time, while neural-network segmentation supports privacy-preserving joint inference with less communication delay.
II. WHERE EDGE INTELLIGENCE MEETS AUTONOMOUS DRIVING: APPLICATIONS AND CHALLENGES
Edge intelligence applies AI-assisted edge computing to autonomous-driving tasks such as object detection, traffic-flow prediction, and path planning. These applications address computation, bandwidth, storage, and latency demands that challenge onboard and cloud-based processing.
- Edge intelligence in autonomous driving supports computation-intensive applications including object detection, traffic-flow prediction, and path planning.
- Real-time video analytics requires timely and accurate object identification, but deep-learning execution can exceed onboard resources and consume substantial bandwidth when tasks are uploaded to remote clouds.
- Edge servers near data sources can receive part of the video analysis and improve inference accuracy for self-driving vehicles.
2) Traffic Flow Prediction:
Autonomous vehicles need timely traffic-flow information and intelligent path planning, but massive environmental data and changing surroundings strain traditional onboard and cloud-based approaches. Edge processing can reduce communication delay and support faster decisions.
- Massive environmental data from multiple sources can saturate onboard storage and exceed traditional computing capabilities, despite the importance of traffic-flow information for navigation applications.
- Online path planning uses AI to support braking, turning, and overtaking decisions while avoiding collisions with surrounding obstacles, vehicles, and infrastructure.
- Edge intelligence decreases wireless communication delay by processing uploaded vehicle-condition and surroundings data near the data sources, enabling timely decisions as conditions change.
- 1) Infeasible Sensing:: Restricted sensor perception and image-quality variation can impair visual object detection, causing a pretrained model to misclassify objects in difficult conditions.
2) Trade-off between Reliability and Latency:
Autonomous-driving inference must balance accuracy and delay under imperfect data, limited edge resources, and privacy-sensitive sensing. Wireless technologies and alternative offloading paths can improve connectivity, but resource allocation and direct data sharing remain central constraints.
- Inference performance depends on both accuracy and delay, while bad-quality data and shallow neural networks can degrade accuracy.
- Future 6G technologies such as mmWave, massive MIMO, and holographic reconfigurable intelligent surfaces may improve wireless-link achievable rates.
- Vehicles may offload data directly to neighboring vehicles through cellular-V2V or PC5 interfaces when shorter device-to-device links are available.
- Edge servers face constrained computation, caching, power, and bandwidth resources, making joint optimization of offloading decisions and resource allocation important.
- Directly sending privacy-sensitive images from autonomous vehicles to edge servers can expose user privacy.
III. MAIN FUNCTIONALITIES IN AUTONOMOUS DRIVING EDGE INTELLIGENCE
Edge intelligence supports autonomous driving through coordinated sensing, distributed learning, and online inference. The functionality modules organize vehicle data collection and model training for connected intelligence.
- Functionality Overview: The four functionality modules are illustrated in Fig. 2 as the organizational structure for edge intelligence in autonomous driving.The figure is referenced as an overview of the system’s functionality modules.
- Data Collection Module: The data collection module gathers sensing data from cameras, LiDAR, and radar alongside vehicle-status data describing onboard computation resources.These inputs support downstream edge-intelligence services for autonomous driving.
- Distributed Learning Module: Distributed learning trains a shared model from local datasets while preserving privacy through repeated parameter uploads and aggregation.The process can be time-consuming under limited communication bandwidth, although B5G/6G may improve its prospects.
C. Offloading Strategy Decision Module
The offloading strategy module addresses time-critical decisions about whether vehicles offload tasks and how edge computation resources are allocated. Online inference complements this process by using local or edge computation while facing privacy and model-size constraints.
- Offloading Strategy: Offloading strategy decisions jointly cover the binary offloading choice and computation-resource allocation, with latency and energy consumption as performance metrics.These decisions are crucial within autonomous-driving safety time windows.
- Online Inference: Local inference is limited by vehicle computing capability, whereas edge inference can provide low-latency services through a trained model deployed at the edge server.The two modes place inference computation in different locations.
- Online Inference: Edge inference still faces data-privacy challenges and limited computation capability for fully bearing and storing sizeable trained ML models.The limitation applies despite the low-latency potential of edge inference.
IV. TWO-TIER EDGE INTELLIGENCE-EMPOWERED AUTONOMOUS DRIVING FRAMEWORK
The proposed framework uses a two-tier architecture for EI-enabled autonomous driving and formulates resource-aware offloading as a difficult optimization problem. An offline-trained multi-task model then enables near-real-time inference of offloading solutions at the base station.
- Two-Tier Framework: The framework divides autonomous driving into an AV-tier and an EI-tier to support EI-empowered object detection.The authors present the architecture with associated design guidelines and an application example.
- Intelligence Offloading: Binary offloading decisions and computational-resource allocation are formulated as a mixed-integer nonlinear programming problem.The formulation is generally NP-Hard, while vehicle counts, computation capability, and channel conditions may vary over time.
- Intelligence Offloading: An offline-trained multi-task neural network learns the mapping from problem inputs to optimal solutions for near-real-time inference at the base station.The base station performs feedforward calculation after receiving vehicle offloading requests.
B. Edge-Vehicle Joint Inference
Edge-vehicle joint inference divides online inference between vehicles and the edge server. Choosing the split point is central because intermediate activations can otherwise amplify communication data rather than reduce it.
- Inference Modes: Online inference has local, edge, and joint modes, with joint inference uploading intermediate parameters to the edge server.Local inference depends on the vehicles’ limited computation resources and storage space.
- Split-Point Selection: Data amplification can make intermediate neural-network outputs larger than the input data, limiting the latency gains of joint inference.The splitting point must therefore be selected so uploaded parameters remain far smaller than raw data.
C. Case Study: EI-Assisted Visual Object Detection
The case study combines MTL-based offloading with segmented edge-vehicle inference for visual object detection, assigning computation across resource-constrained AVs and an edge server. This design supports near-real-time decisions while improving privacy and adapting inference to image quality.
- Multi-task Learning Model: The MTL model formulates binary offloading as classification and edge-server resource allocation as regression for MINLP solutions.The model uses cross-entropy and MSE losses, respectively, and has a size below 2 KB.
- Enhanced-inference results sent to ECU: The framework’s case study sends enhanced inference results to the vehicle ECU after processing across the vehicle and edge tiers.The workflow is presented as a two-tier autonomous driving framework.
- Offloading strategy inferred by MTL model: The edge server infers offloading strategies through feedforward calculation without iterations, shifting online complexity to offline training.The method is intended to provide high-accuracy decisions in near-real-time.
- Edge-Vehicle Joint Inference: Neural network segmentation places shallow SNN layers on AVs and deeper DNN layers at the edge server for privacy-preserving joint inference.All AVs share the edge-hosted DNN layers, while vehicles retain limited local computation and storage requirements.
- Edge-Vehicle Joint Inference: Low-quality images can trigger offloading of intermediate feature maps to the edge, whose deeper layers return results to help vehicles react accurately.This design uses edge processing to improve inference accuracy while avoiding direct transfer of the original captured images.
D. Illustrative Results
The experiments evaluate training-data effects, compare MTL with sBB for offloading decisions, and examine inference strategies under varying bad-data ratios. MTL improves solution quality and speed, while edge inference becomes more advantageous as image quality deteriorates.
- Training-sample impact: Inference accuracy for binary offloading increases with training-sample percentage, while regression MSE decreases as training data grows.The evaluation uses 4 × 10^4 generated samples and sets χc = χr = 1.
- MTL versus sBB: MTL always achieves higher inference accuracy than sBB, with almost 120% improvement when the number of AVs is 8.The MTL regression MSE is also less than half that of sBB.
- MTL versus sBB: Using multi-task learning, MINLP-solving time is less than one-thousandth of the sBB scheme for 2 to 8 AVs.This comparison concerns the time required to solve the offloading optimization problem.
- Inference under bad data: When the bad-data ratio η is below 0.3, vehicle inference is more competitive because it avoids communication delay.As η reaches or exceeds 0.3, poor SNN inference causes a sudden weighted-sum-cost increase.
- Inference under bad data: For η ≥ 0.3, edge inference benefits from deeper DNN layers that improve accuracy through offloading intermediate feature maps.The weighted-sum cost increases with η across the three evaluated inference methods.
V. CONCLUSION AND OPEN RESEARCH TOPICS
The article presents a two-tier edge intelligence architecture for connected autonomous driving and identifies open challenges in data, bandwidth, sensing, and vehicle collaboration.
- The proposed two-tier architecture combines edge learning with functionality designs for efficient offloading, data privacy, improved inference accuracy, and delay constraints.Its effectiveness is demonstrated through a case study and experiments.
- Efficient edge learning with limited labeled training data: Limited labeled training data constrains supervised edge learning, motivating historical-data reuse and semi-supervised, transfer, or autonomous learning.Sensor fusion with LiDAR and radar is also proposed to address deficiencies from video-only sensing.
- Distributed model training with restricted wireless bandwidth: Restricted wireless bandwidth makes distributed model training costly because model exchanges are communication-intensive and iterations are limited by the slowest participants.Federated learning preserves privacy but faces non-independent data and runtime bottlenecks across autonomous vehicles.
- Vehicle platoon-aided inference: Vehicle platoon-aided inference can maintain reasonable inference outside base-station coverage by routing intermediate parameters among vehicles.The approach trades improved inference accuracy for V2V communication delay, while platoon dynamics remain an open issue.