PaperlayerView exact revision on arXiv

Source-linked AI summary

Decision-Making in Driver-Automation Shared Control: A Review and Perspectives

Wenshuo Wang, Xiaoxiang Na, Dongpu Cao, Jianwei Gong, Junqiang Xi, Yang Xi, Fei-Yue Wang

arXiv:2007.12597v1eess.SY

TL;DR

Driver-automation shared control must determine how human drivers and automated agents share roles, authority, and decision-making while preserving useful human involvement. This paper surveys shared-control architectures, driver models, and interaction strategies, then discusses challenges including stochastic human behavior, personalization, authority allocation, and transition design.

  • Problem

    Shared-control design must address how to assign human roles, authority, responsibility, trust, and adaptation while accounting for complex and time-varying human decision-making.

  • Method

    The paper surveys architectures, decision-making schemes, human driver models, and interaction strategies in driver-vehicle shared control.

  • Results

    The review organizes the field into shared-control architectures and modeling approaches, and identifies future challenges involving stochastic human behavior, personalization, authority allocation, and human–automation transitions.

  • Takeaways & Limitations

    A deeper and more holistic understanding of relationships between human drivers and automated driving agents remains important for developing driver-vehicle shared-control systems.

Abstract

from arXiv · show

Shared control schemes allow a human driver to work with an automated driving agent in driver-vehicle systems while retaining the driver's abilities to control. The human driver, as an essential agent in the driver-vehicle shared control systems, should be precisely modeled regarding their cognitive processes, control strategies, and decision-making processes. The interactive strategy design between drivers and automated driving agents brings an excellent challenge for human-centric driver assistance systems due to the inherent characteristics of humans. Many open-ended questions arise, such as what proper role of human drivers should act in a shared control scheme? How to make an intelligent decision capable of balancing the benefits of agents in shared control systems? Due to the advent of these attentions and questions, it is desirable to present a survey on the decision-making between human drivers and highly automated vehicles, to understand their architectures, human driver modeling, and interaction strategies under the driver-vehicle shared schemes. Finally, we give a further discussion on the key future challenges and opportunities. They are likely to shape new potential research directions.

I. INTRODUCTION

Driver-vehicle shared control is presented as a transitional framework for combining human and automated capabilities while addressing authority transfer, human involvement, and decision-making challenges. The review surveys these issues and organizes open research questions around roles, responsibility, trust, adaptation, and driver-state identification.

  • Motivation: Level 3 automation requires transitions between automated and manual control, but resuming control can degrade performance when drivers are not actively engaged.Shared control is proposed as a way to support continuous authority transition between the human driver and automated driving agent.
  • Motivation: Shared control combines human and automated agents so each can contribute its strengths to driving tasks while compensating for human errors and automation limitations.Manual control is vulnerable to human error, whereas fully automated tasks remain limited in decision-making and situation awareness.
  • Scope: Driver-vehicle shared control includes perception enhancement, action suggestion, and function delegation, with simultaneous human and automated inputs exemplified by active steering assistance.The reviewed focus is the function-delegation form in which both agents can exert vehicle inputs during a task.
  • Research questions: The review identifies open questions about human roles, authority and responsibility allocation, mutual trust evaluation, adaptation and learning, and nonlinear driver-state and intent identification.These questions concern changing situations and the respective abilities of human drivers and automated driving agents.
  • Contribution: The paper reviews more than 200 related publications to survey decision-making scheme design and human driver modeling in shared control systems.Its emphasis is scheme design rather than rigorous mathematical algorithms for decision-making and controller design.

II. SHARED CONTROL ARCHITECTURES

The paper distinguishes task-level and servo-level shared control architectures according to how human and automated agents divide driving authority and inputs. Servo-level schemes combine their outputs directly or indirectly, with the weighting coefficient governing the human–automation contribution.

  • Architecture: A driver-vehicle shared-control system is a two-agent system that accepts and executes commands from the human driver, the automated driving agent, or both.The architecture is introduced before reviewing decision-making and driver-modeling approaches.
  • Servo-Level Shared Control: Servo-level shared control keeps the human driver engaged in vehicle movement while combining human operation output uh with automated output uc.Unlike task-level control, the automated agent does not simply take over task-specific control at the servo level.
  • Task-Level Shared Control: Task-level shared control lets the human driver decompose a driving task into subtasks and allocate selected subtasks to the automated driving agent.The automated agent then performs assigned subtasks using current situation conditions and predefined algorithms.
  • Servo-Level Shared Control: The weighting coefficient λ ∈[0, 1] determines the proportion of human and automated inputs in the vehicle command, ranging from pure human control at λ = 1 to pure automation at λ = 0.The coefficient may be fixed, manually adjusted, or continuously adapted according to driver state.
  • Servo-Level Shared Control: The servo-level architecture includes direct shared control, where both agents act simultaneously through a control interface, and indirect shared control, where inputs are mixed before reaching the vehicle.Direct schemes are commonly haptic, whereas indirect schemes can use steer-by-wire systems.

1) Direct shared control:

Driver-in-the-loop shared-control models represent vehicle dynamics, human operation, and automated control together to support analysis and decision-making. The review distinguishes direct and indirect control interfaces and describes state-space and optimization-based formulations.

  • Direct shared control: Direct shared control lets human drivers and automated agents simultaneously act on a haptic interface, with both influencing vehicle inputs.Typical interfaces include the steering wheel and brake or throttle pedals, allowing drivers to perceive assistance torque.
  • Indirect shared control: Indirect shared control mixes human contributions and automated-agent outputs before shaping the controlled vehicle input.Steer-by-wire systems can estimate the driver’s desired steering angle and apply it directly to the front wheels.
  • Driver-in-the-loop modeling: State-space driver-in-the-loop models combine road-vehicle dynamics with human operation and controller inputs, including disturbances.The augmented state contains both driver-model and road-vehicle states.
  • Driver-in-the-loop modeling: Driver models can be formulated from control, stochastic, or cognitive perspectives, with preview models commonly integrated into vehicle state-space models.Drivers compare desired and current trajectories through an internal model that estimates vehicle and surrounding-object states.
  • Decision-making: Optimization-based decision-making can include driver input, vehicle dynamics, and physical limitations while requiring a formulation of control-authority allocation.The allocation relationship among automated, human, and combined inputs is represented through the authority coefficient λ.

A. Rule-Based

Rule-based decision-making designs the shared-control authority coefficient λ(t) using predefined rules derived from prior knowledge.

  • A. Rule-Based: Rule-based methods design λ(t) according to prior knowledge, with existing rules grouped into three categories.The approach directly uses the shared-control formulation while specifying how authority allocation changes over time.

1) Piecewise Function:

Piecewise and exponential functions provide rule-based ways to adapt shared-control authority to driver behavior, driving situations, and safety needs. U-shape models instead estimate assistance from the relationship among workload, performance, and assistance need.

  • 1) Piecewise Function: Piecewise functions divide operating conditions into regions and assign different shared-control laws to each region.Their robustness and practicality motivated their adoption from robotics into intelligent vehicles.
  • 1) Piecewise Function: When the human behaves dangerously, automated feedback resumes authority; safe behavior leaves control to the human; hysteresis uses a predefined shared law.This rule explicitly maps behavioral regions to authority outcomes.
  • 1) Piecewise Function: Piecewise functions can incorporate human factors and driving situations using prior knowledge, while combined piecewise schemes can assess safety and vehicle performance.Other designs use human-applied steering torque to estimate the assistance torque delivered by the agent.
  • 2) Exponential Function: Exponential functions produce continuous shared-control factors by combining an activity factor with tuning and safety parameters.The safety parameter is intended to guarantee safe operation and model convergence.
  • 3) U-Shape Function: U-shape functions estimate assistance from driver activity and workload-performance relationships, with applications including lane keeping and obstacle avoidance.One formulation computes assistance torque as Ta = µ(a)Ts, where driver activity a is represented by steering angle.

B. Game-Theory-Based

Game-theoretic shared-control models treat the human driver and automated driving agent as dynamically interacting decision-makers. The review distinguishes noncooperative and cooperative formulations, including Nash, Stackelberg, and Pareto schemes, while noting that incorrect state or intent estimates can undermine cooperation.

  • Game-theory-based shared control: Game-theoretic approaches model the human driver and automated driving agent as two interacting players whose actions influence one another.Their objective functions depend on vehicle state and both agents’ inputs.
  • Game-theory-based shared control: Dynamic games are classified into noncooperative and cooperative forms according to how the driver and automated agent interact.Noncooperative agents pursue individual interests, whereas cooperative agents seek binding agreements and global optimality.
  • Game-theory-based shared control: Nash strategies are simultaneous and mutually responsive, while Stackelberg strategies are sequential and assign leader-follower roles.These assumptions produce different game-theoretic schemes for shared-control decision-making.
  • Game-theory-based shared control: Cooperative Pareto schemes account for both agents’ desired trajectories and actions, with Nash or Stackelberg equilibria used to solve coupled optimization problems.Solutions may be analytical or approximated, and prior work includes continuous role adaptation.
  • Game-theory-based shared control: Incorrect estimates of driving situations or driver intent can prevent the human and automated agent from cooperating.Poorly designed shared-control strategies may also cause loss of expertise, complacency, trust, and adaptivity.

IV. HUMAN DRIVER MODELING

Effective driver-vehicle shared control requires modeling human drivers alongside vehicle dynamics. The review emphasizes sensory dynamics, cognition, hidden states, and operational characteristics as central aspects of human-driver modeling.

  • IV. HUMAN DRIVER MODELING: Human-driver models should represent sensory dynamics, cognition processes, hidden states, and operation characteristics alongside vehicle dynamics.These elements support understanding whether humans and automated vehicles correctly interpret each other’s intentions.

A. Functional Module

Driver sensory dynamics provide important inputs for vehicle speed and direction control and can support safer, more compatible driver-vehicle system design. Visual information is especially important, while vestibular, kinesthetic, somatosensory, and auditory cues supplement perception.

  • A. Functional Module: Driver sensory dynamics are important for designing driver-vehicle systems that interact more safely and naturally.The review ranks visual, vestibular and kinesthetic, somatosensory, and auditory channels for regular driving tasks.
  • A. Functional Module: Visual information is the most significant sensory source in normal driving and accounts for about 90% of sensory information.Visual cues include information about the future road trajectory.
  • A. Functional Module: Driver visual information can indicate lane-change intentions, readiness to take over automation, fatigue, and upcoming actions.Eye gaze is one example of a visual signal used to infer these latent states and intentions.
  • A. Functional Module: Vestibular and kinesthetic cues supplement visual information by providing vehicle-motion information such as acceleration and rotation.These channels also contribute to combined arm-trunk motion during control.
  • A. Functional Module: Somatosensory inputs include pedal pressure and steering-wheel torque, while auditory information typically serves as a supplementary cue.Sensory channels have physical characteristics including time delay, perception limits, and coherence zones.

2) Cognition:

Cognitive driver modeling addresses discrete driving decisions, distraction, and the coordination of assistance with human information processing. The review also highlights neuromuscular dynamics and limitations of control-theoretic models that omit cognition and decision-making.

  • 2) Cognition:: Human-driver cognition governs discrete decisions such as whether, when, and how to change lanes, motivating tactical and strategic behavior models.These models address high-level behaviors including speed selection and decision-making.
  • 2) Cognition:: Cognitive models have been used to combine experimental knowledge, detect distraction, and analyze how cognitive and visual distractions affect vehicle performance.One real-time distraction model considered distraction definition, model inputs, and input summarization.
  • 2) Cognition:: Cognitive assistance can determine when steering assistance should be provided and how much support to deliver.The proposed assistance stages address acquiring information, analyzing information, and making decisions.
  • 2) Cognition:: Neuromuscular-skeletal dynamics constrain perception-action coupling and matter for avoiding subtle conflicts between human drivers and automated agents.The review focuses on arm dynamics for lateral control and foot dynamics for longitudinal control.
  • 2) Cognition:: Single- and two-point visual-preview models are easy to integrate with vehicle models but assume known desired trajectories and omit cognitive characteristics such as neuromuscular dynamics and decision-making.Control-theoretic models also commonly overlook how drivers perceive external inputs.
  • 2) Cognition:: Game-theoretic shared control requires models of both the human driver and automated agent so each can estimate the other’s actions.Control theory, including model predictive control, is a common approach for modeling automated agents’ predictions of future vehicle states.

2) Learning-Based Driver Model:

Learning-based driver models extend behavior prediction to dynamic and stochastic intent and action processes, using methods such as Markov, Bayesian, and deep-learning approaches. These models support intent inference and shared-control design through haptic, neuromuscular, observational, and environmental cues.

  • Learning-Based Driver Model: Existing driver models often predict behavior, actions, and states but do not represent drivers’ dynamic, stochastic decision-making processes.The review identifies a need to connect temporal and spatial processes in driver decision models.
  • Learning-Based Driver Model: Learning-based approaches include neural networks, Markov models with Gaussian mixtures, Bayesian inference, ARX models, and deep learning.These approaches address the nonlinear and temporally structured properties of driver behavior.
  • Learning-Based Driver Model: Steering-wheel torque provides a direct, continuous interface for capturing driver intent while human drivers and automated agents act simultaneously.Torque-based assistance systems use sensors to measure the driver’s steering input and exchange information through the steering interface.
  • Learning-Based Driver Model: Neuromuscular models of driver arms and legs can guide shared-control design for steering, accelerator, and brake systems.Combining driver neuromuscular behavior with steering dynamics supports models that reflect interaction between the driver and automation.
  • Learning-Based Driver Model: Changes in driver-arm damping and stiffness can indicate whether controller outputs satisfy the desired trajectory or reflect conflict with the automated agent.The model relates inertia, damping, stiffness, muscle torque, steering ratio, and steering angle to driver intent inference.
  • Learning-Based Driver Model: Driver intent can also be inferred from body gestures, gaze, head, hands, feet, peripheral vehicles, and vehicle position relative to lane edges.Dynamic Bayesian networks, Markov decision processes, and partially observable Markov decision processes are identified as relevant modeling frameworks.

2) State Detection:

State detection supports authority allocation and shared-control design by estimating driver sleepiness, distraction, intent, and other operating conditions. The review also emphasizes unresolved questions about adaptive roles, trust, function delegation, and personalization.

  • State Detection: Detecting sleepiness, drowsiness, fatigue, distraction, and impairment can support practical authority allocation decisions.One lane-keeping assistance scheme detects sleepiness using eye-blinking frequency and facial information.
  • State Detection: Visual and cognitive distraction detection combines vehicle state, visual state, and driver operations using deep autoencoders, DBNs, and SVMs.These learning-based methods have been used to detect and classify driver distraction.
  • State Detection: Open questions concern adaptive laws, appropriate trust, human-driver roles, and how to implement function delegation appropriately.The review states that systematic answers remain unavailable while research commonly examines haptic support, vehicle stability, and collision avoidance.
  • State Detection: Shared control performance can be improved by considering individual characteristics, classifying human-automation interaction levels, and integrating cognitive psychology.The review presents these approaches as methods for adapting shared control to drivers and interaction conditions.
  • State Detection: The U-shape control law qualitatively relates workload, assistance needs, and driving performance but lacks a quantitative expression.Differences among derived adaptation laws may reflect variation in drivers’ experience and physical or psychological status.
  • State Detection: Future shared-control research must account for stochastic driver behavior, differences among drivers, and impaired states rather than assuming identical deterministic objectives.The review also identifies the human driver’s role relative to the automated agent as unresolved.

3. Which way is the best transition in authority, ability, responsibility, and control (A2RC)?

The review frames authority, ability, responsibility, and control as interrelated dimensions of human-vehicle shared control. It highlights unresolved allocation questions and examines cognitive, physiological, and interface-based approaches to safer authority transitions.

  • 3. Which way is the best transition in authority, ability, responsibility, and control (A2RC)?: Authority can be partially and continuously transferred between the human driver and automated agent.It concerns what each agent is allowed to do or not do.
  • 3. Which way is the best transition in authority, ability, responsibility, and control (A2RC)?: Ability concerns the means or skill to perceive, select adequate actions, and act appropriately.The definition treats ability as a capability dimension in the A2RC framework.
  • 3. Which way is the best transition in authority, ability, responsibility, and control (A2RC)?: Responsibility concerns prior assignment of accountability and subsequent evaluation of who is accountable or to blame for system states, actions, and consequences.Control separately refers to the power to influence vehicle states.
  • 3. Which way is the best transition in authority, ability, responsibility, and control (A2RC)?: The review asks how to balance powerful automation with retained human authority and how to assign ability, authority, and responsibility across future system subsystems.It notes that much literature focuses on control authority while overlooking relationships among the four A2RC concepts.
  • 3. Which way is the best transition in authority, ability, responsibility, and control (A2RC)?: Brain-related signals and psychologically informed interfaces may provide information about driver intent and ability for authority allocation and conflict reduction.The review also points to human psychological, psychobiological, and active capabilities as interface-design considerations.
  • 3. Which way is the best transition in authority, ability, responsibility, and control (A2RC)?: Human factors influence automation-to-manual transitions, but the effects of torque, steering angle, and vibration on shared-control performance remain unclear.The review identifies this uncertainty as a continuing research issue in transition-interface design.

VI. DISCUSSION AND CONCLUSION

The review surveys architectures, modeling approaches, and interaction strategies for driver-vehicle shared control while identifying future challenges and opportunities. It concludes that holistic understanding of driver-agent relationships and cognitive interaction with the driving environment remains an active research area.

  • VI. DISCUSSION AND CONCLUSION: The paper reviews driver-vehicle shared-control architectures, complex-system modeling approaches, and future challenges and opportunities.It surveys progress in shared-control technologies over the past decades.
  • VI. DISCUSSION AND CONCLUSION: The review decouples complex driver-vehicle systems into subsystems and summarizes state-of-the-art approaches for modeling them.This decomposition organizes the survey of shared-control technologies and driver-vehicle relationships.
  • VI. DISCUSSION AND CONCLUSION: A deeper and more holistic understanding of human-driver relationships with automated agents and cognitive interaction with the driving environment remains an active research area.The paper places this issue among the continuing challenges following advances in driver assistance systems.
Loading 2007.12597v1…