The problem
Lane keeping systems that fight the driver are annoying and can be unsafe, while systems that hand over abruptly create confusion. Haptic shared control lets driver and automation steer together, but how much authority the automation should take, and how to detect driver intent quickly, remain open questions with direct impact on safety and acceptance.
Objectives
- Model the steering system, vehicle and a cybernetic driver model in Simulink.
- Design a shared control law that blends automation torque with driver torque.
- Develop an authority allocation strategy based on driver intent and risk.
- Evaluate conflict, tracking performance and safety with driver models and human participants.
- Compare fixed, intent-adaptive and risk-adaptive authority strategies.
What you need
Tools and toolboxes
- MATLAB — Analysis of experiments and statistics.
- Simulink — Steering column, vehicle dynamics, driver models and the shared controller.
- Model Predictive Control Toolbox — An MPC-based automation layer that includes driver torque as a measured disturbance.
- Automated Driving Toolbox — Scenarios with roads, obstacles and lane boundaries, and its Unreal Engine 3D simulation blocks for the driver display.
- Simulink Desktop Real-Time (optional) — Running the model in real time with a force-feedback wheel.
Software
- MATLAB and Simulink
- Model Predictive Control Toolbox
- Automated Driving Toolbox
- Simulink Desktop Real-Time or a similar real-time I/O solution
Extra parts and data
- Force-feedback gaming steering wheel and pedals — Needed for driver-in-the-loop experiments so participants feel and apply steering torque.
- Institutional ethics approval for participant studies — Human-subject experiments need informed consent and approval before data collection.
Key concepts
- Haptic shared control
- The automation applies a guiding torque that the driver feels and can overrule, so both continuously influence the steering rather than switching between modes.
- Level of haptic authority
- How strongly the automation's torque acts relative to the driver's, set by the controller gain or stiffness; high authority resists the driver more.
- Driver intent inference
- Estimating from torque, steering angle, gaze or indicator use whether the driver wants a different path, so the automation can yield.
- Conflict torque
- The part of driver and automation torques that oppose each other; a key metric of how much they disagree.
Step-by-step guide
- 1
Frame the research questions
Review haptic shared control, authority allocation and driver modelling literature. Pose concrete questions, for example whether risk-adaptive authority reduces conflict torque without increasing lane departures compared with fixed authority. Define the ODD: highway and rural curves with occasional obstacles at moderate speed. List metrics: lateral error, conflict torque, driver steering effort, time to resolve conflicts, minimum distance to obstacles and subjective workload.
Checkpoint: A research proposal chapter with questions, hypotheses and a metric list.
- 2
Model steering and vehicle
Build a Simulink steering column model with wheel inertia, damping, stiffness, a torque sensor and an assist motor, coupled through a rack to a dynamic bicycle model. Parameterise it from published data for a passenger car with electric power steering. Validate the steering feel by checking the torque-angle relationship at different speeds and the vehicle's yaw response to steering steps.
Checkpoint: Steering torque against angle and yaw responses that match expected trends.
- 3
Model the driver
Implement a cybernetic driver model, such as a two-point visual model with near and far points and a neuromuscular arm model with admittance and reflex delay. Tune it to produce realistic lane keeping without automation. Create variants: an attentive driver, a distracted driver with delayed response and a driver who intends a different path, for example to avoid an obstacle the automation has not seen.
Checkpoint: The driver model alone keeps lane on test roads with plausible torque levels and reaction times.
- 4
Design the automation layer
Design an MPC or a look-ahead controller that computes a desired steering angle for lane keeping and obstacle avoidance, then converts it into a guidance torque through an adjustable stiffness. Include driver torque as a measured input so the automation predicts the combined effect. Ensure the automation torque is bounded so a driver can always overrule it with moderate effort.
Checkpoint: With a passive driver, the automation alone keeps lane and avoids obstacles in simulation.
- 5
Design authority allocation
Implement three strategies: fixed authority; intent-adaptive, which lowers stiffness when sustained driver torque opposes the guidance; and risk-adaptive, which raises authority when time to lane crossing or time to collision drops. Smooth all changes in authority to avoid torque steps. Define intent detection features and thresholds, and log authority over time. Choose the intent features from the driver-model runs, and check that ordinary corrective steering by an attentive driver is not mistaken for a wish to leave the automation's path.
Checkpoint: Each strategy shows the expected authority changes in scripted test cases.
- 6
Evaluate with driver models
Run the three strategies with each driver model across a scenario set of curves, lane drift during distraction and an obstacle visible only to the driver. Use randomised driver parameters for repeated runs. Compute all metrics and compare strategies statistically. Identify failure modes such as automation resisting a correct avoidance manoeuvre. Use the same seeds for every strategy so differences come from the allocation law, and plot authority, driver torque and lateral position together for the most informative runs.
Checkpoint: A simulation results table with mean and spread for each strategy and driver type.
- 7
Set up the driver-in-the-loop simulator
Connect a force-feedback wheel to the Simulink model in real time, mapping measured wheel angle and driver torque into the model and commanding the wheel's force feedback. Render the road scene on a monitor. Check end-to-end latency and torque fidelity. Prepare a short familiarisation drive. Obtain ethics approval and design the experiment: counterbalanced strategy order and a set number of participants.
Checkpoint: A working simulator with measured latency and an approved experiment protocol.
- 8
Run participant experiments
Run each participant through the scenarios with each strategy, recording steering angle, torques, lateral position, events and gaze if an eye tracker is available. After each condition, collect workload and acceptance questionnaires such as NASA-TLX and the Van der Laan scale. Keep the sessions short to limit fatigue and simulator sickness, and allow participants to stop at any time. Note any session stopped early and the reason.
Checkpoint: Complete datasets for all participants and conditions with questionnaire scores.
- 9
Analyse the results
Use repeated-measures statistics or mixed-effects models to compare strategies on objective metrics and questionnaire scores. Check whether simulation predictions with driver models agree with participant results, and where they differ. Relate conflict torque to subjective acceptance. Report effect sizes and confidence intervals, not just p-values. Check model assumptions such as normality of residuals, and use non-parametric tests where they fail. Report how many participants showed each effect, since group averages can hide individuals for whom a strategy worked badly.
Checkpoint: Statistical results with effect sizes and a comparison between model and human data.
- 10
Write up and position the work
Write the thesis chapters covering models, controller, allocation strategies, simulation study and human experiments. Discuss implications for ISO 26262 and SOTIF analysis, such as the hazard of automation resisting a correct driver action. State limitations: simulator realism, participant sample and scenario coverage. Identify what a closed-track study would add in future. Prepare a short paper that focuses on the comparison of allocation strategies, with the models and experiment protocol described in enough detail to repeat the study.
Checkpoint: Thesis chapters and a draft paper with clear contribution statements.
Testing & validation
- Lateral error and lane departures per strategy and driver type.
- Conflict torque and driver effort.
- Response time when driver intent differs from automation.
- Subjective workload and acceptance from participant questionnaires.
- Agreement between driver-model simulations and participant results.
Expected results
A shared steering controller with three authority strategies, a validated simulation framework with driver models, and a participant study quantifying conflict, safety and acceptance trade-offs.
Research angle for your PhD
Research questions: how should haptic authority be allocated between driver and automation to minimise conflict while preserving safety, and can driver-model simulations predict human outcomes? Metrics: conflict torque, lateral error, minimum TTC, takeover response time, NASA-TLX and acceptance scores. Contributions can include a risk- and intent-adaptive allocation law, a validated simulation-plus-participant evaluation method and evidence on model-human agreement, which are suitable for journal and conference submission.
Take it further
- Add gaze-based intent detection with an eye tracker.
- Extend to takeover requests from higher automation levels.
- Study two-wheeler or heavy-vehicle variants.
Viva and review questions
What is the main risk of high haptic authority?
The automation may resist a correct driver manoeuvre, for example avoiding an obstacle the sensors missed, increasing conflict and possibly delaying the avoidance.
How do you detect driver intent?
From sustained driver torque opposing the guidance, steering angle deviation and indicator use, filtered over a short window to avoid reacting to noise.
Why combine driver models with human experiments?
Models allow many repeatable runs and early tuning, while human studies capture real variability and acceptance; comparing both tests the models themselves.
How did you handle ethics?
Participants gave informed consent under an approved protocol, sessions were short with the option to stop, and data was anonymised.
Learn it at DIYguru
These DIYguru courses and programmes teach what this project needs.