Haptics Based Multi-Level Collaborative Steering Control for Automated Driving

The paper addresses how to achieve safe, acceptable partial automation by keeping drivers actively engaged through collaborative steering. In Level-2 automation, vehicles handle sustained lateral/longitudinal control while the driver remains responsible for object and event detection and response. Existing haptic shared control (HSC) implementations commonly use blended (parallel force/position) steering control, which struggles to balance high trajectory tracking with allowing manual intervention, often leading to discontinuous ADAS behavior (e.g., overrides in lane centering and lane change, separate LCA/LKA functions, and uncomfortable switching). The research proposes a driver-oriented automation strategy that maintains shared control continuously and integrates ADAS functions consistently. The framework centers on three functions: interaction (enabling haptic deviation without harming tracking via admittance control), arbitration (allocating authority based on interaction type), and inclusion (assimilating persistent driver intent into trajectory planning). The goal is to improve trust, comfort, and safety by enabling intuitive, continuous collaboration across operational and tactical levels.

Prior work on shared steering control is extensive (>100 studies) but often focuses on isolated issues such as priority and conflict management, lacking a unified approach. Blended control’s dual role (tracking and disturbance rejection) forces gain compromises, resulting in limited tracking in curves, override thresholds, and discontinuous operation across LCA, ALC, and LKA. Studies addressing interaction and arbitration alone (e.g., fixed or modulated impedances) lack inclusion, so the system reverts to the nominal trajectory after driver input. Conversely, approaches that adapt trajectories from human force/torque inputs (inclusion without arbitration) act at vehicle-motion bandwidths, unsuitable for haptic interaction quality. The paper positions its contribution as integrating interaction, arbitration, and inclusion simultaneously within mass-production constraints, extending beyond previous frameworks by enabling continuous shared control, driver-initiated rerouting, and consistent actuator coordination.

Control framework: A multi-level collaborative steering framework integrates three functions: (1) Interaction via admittance control to enable manual deviation while keeping high tracking performance; (2) Arbitration to allocate automation authority based on estimated driver motor control and a preset interaction type (co-activity, collaboration, competition); (3) Inclusion to assimilate driver intent into trajectory planning for consistent multi-actuator behavior. System dynamics: The steering system is modeled as a dual-pinion EPS with steering wheel and pinion inertias. Dynamics include motor torque command, driver torque, torque sensor stiffness, and disturbances (friction, backlash, road load). Both driver and automation are modeled as impedance controllers with goals (target steering angles) and impedances. Time-varying impedance dynamics are represented with first-order models. Interactive steering control (admittance): A stiff inner angle control loop provides high tracking accuracy of the AD trajectory, while an outer torque loop closes only when driver torque is applied. A virtual plant estimates manual deviation θ_m from measured driver torque and system dynamics, and the inner loop enforces the superposed command θ_cmd = θ_a + θ_m. The equivalent interaction dynamics form a two-inertia system; stability is ensured by setting the outer loop bandwidth below the inner loop and by selecting virtual inertia larger than the actual steering wheel inertia. Arbitration: Automation impedance Z_o is adapted from its nominal value according to estimated driver impedance Z_d and parameter κ setting interaction type: Z_o = Z_o − κ Z_d. κ = 0 yields co-activity (constant automation impedance), κ > 0 yields collaboration (automation impedance decreases as driver impedance rises), and κ < 0 yields competition (automation impedance increases with driver impedance). Cooperation/assistance is excluded due to independent agent goals. Driver motor control estimation: To avoid observability issues of jointly estimating driver goal and impedance, the driver goal is roughly approximated first, then impedance estimated with an EKF. The goal combines an environmental constraint (lane center tracking from vehicle and road curvature model) with driver intent estimated via a simple admittance model propagating measured driver torque forward in time. Using measured torque and pinion angle, an EKF (with discretized plant and tuned noise covariances) estimates time-varying driver stiffness and damping. This approach uses only sensors available in mass-produced vehicles. Inclusion into trajectory planning: Manual deviation is converted to vehicle yaw rate via a single-track model, then to desired lateral deviation with a constant turn-rate-and-velocity (CTRV) model over a prediction horizon. The trajectory planner’s cost function is augmented with a term reflecting driver desired lateral position, alongside comfort (jerk), time, and automation lateral error terms. During driver intervention, optimal lateral trajectories are recomputed at a higher rate to assimilate intent; during no intervention, the selected trajectory is tracked. This yields bounded interaction torque and propagates the manual correction consistently to other actuators (e.g., brakes, acceleration). Experimental setups: Four configurations validate components and the integrated framework: (1) Virtual driver bench (EPS + impedance-controlled motor in place of a human) to validate driver goal and impedance estimation; (2) Human driver bench to validate driver impedance estimation and arbitration rules by executing sine/slalom maneuvers under preset interaction types; (3) Static driving simulator to validate trajectory adaptation in a double lane change at 60 km/h using Stanley controller; (4) Test vehicle with high-precision GNSS tracking a nominal AD trajectory at 60 km/h on a three-lane, 1.5 km course performing double lane changes at 100 m intervals. Quantitative evaluation uses KPIs: driver effort (integral of torque over time) and steering entropy (based on prediction-error distribution of steering angle).

Estimation performance (virtual driver): The EKF-based approach approximates driver goal and captures driver impedance variations. Using true impedance with known goal yields oscillatory convergence due to undefined impedance when torque or error tends to zero; combined estimation (approximated goal feeding EKF) overestimates impedance in steady state and amplifies oscillations because of model mismatch, yet preserves useful impedance dynamics for arbitration. Arbitration verification (human driver bench): Switching interaction types every 15 s shows expected behavior. Co-activity (κ=0) yields tracking near the average of driver and automation. Collaboration (κ=1) adaptively reduces automation impedance as driver impedance increases, smoothly transferring authority to the driver during engagement and restoring it when hands-off. Competition (κ<0) increases automation impedance with driver impedance, opposing manual intervention and requiring higher driver torque for the same maneuver. Hands-off periods naturally return authority to automation. Trajectory adaptation (driving simulator): Without inclusion, sustained driver torque is required to maintain lane change against a fixed AD trajectory, providing guidance but causing effort. With inclusion active (κ=0), the AD trajectory shifts to the driver-selected lane when deviation is sufficient; interaction torque remains bounded, reducing sustained effort and centering the vehicle in the new lane. Integrated proof-of-concept (test vehicle): With collaboration (κ=1), complementary impedances reduce torque peaks relative to co-activity and the induced lateral deviation is assimilated into trajectory planning, allowing the vehicle to track new lanes without sustained driver torque. With competition (κ negative), the automation resists deviation; manual intervention is not assimilated and the vehicle follows the original AD trajectory. Driver quantitative study (n=5, age 29–44, mean 35, average annual driving 5800 km): Participants executed double lane changes at 60 km/h under four modes: (1) co-activity without inclusion, (2) collaboration without inclusion, (3) co-activity with inclusion, (4) collaboration with inclusion. Results based on normalized driver effort (DrE) and steering entropy (StE) indicate that modes with inclusion (3, 4) reduce both DrE and StE and show lower inter-participant variability than modes without inclusion (1, 2). Two behavioral groups emerged under modes 1 and 2 (high-effort smooth vs. lower-effort less smooth), suggesting lower acceptance compared to modes 3 and 4. Overall, arbitration plus inclusion (mode 4) delivered smoother maneuvers with less effort across drivers.

The findings demonstrate that integrating interaction, arbitration, and inclusion enables true collaborative steering: drivers can deviate from the AD trajectory with appropriate haptic cues, and persistent intent is assimilated so the vehicle does not revert to the nominal path after driver input. Compared to interaction+arbitration-only approaches, inclusion prevents undesired return to the original trajectory and supports continuous shared operation during lane or route changes. Compared to interaction+inclusion without arbitration, competition mode enables full rejection of driver input when safety or higher automation demands it, accommodating a wide range of automation levels and multi-objective ADAS functions (e.g., temporarily high reaction torque for LKA collision avoidance). Admittance control resolves the blended-control trade-off between tracking accuracy and manual acceptance, widening the tuning range. Quantitative results (DrE and StE) support that arbitration plus inclusion improves smoothness and reduces effort across diverse drivers, enhancing acceptance. The framework’s multi-level integration ensures consistent coordination across actuators and improves haptic consistency by propagating manual deviation into trajectory planning.

The paper proposes a driver-centered multi-level control framework for collaborative steering that operates within mass-production hardware constraints. Using admittance-based interaction, impedance-based arbitration with preset interaction types, and inclusion of driver intent in trajectory planning, the framework enables continuous shared control, consistent ADAS integration, and compatibility with automation levels where the driver remains involved. Practical contributions include: robust high-performance angle tracking with manual deviation capability; a simple arbitration rule enabling a spectrum from competition to collaboration; a pragmatic approximation of driver goal and EKF-based impedance estimation; and propagation of manual deviation to trajectory planning to coordinate all actuators. Vehicle tests with five participants suggest the framework can yield smoother maneuvers with less driver effort and better acceptance. Future work includes selecting interaction types via higher-level situational assessment (driver state, road/traffic), extending arbitration to optimization-based control (e.g., nonlinear MPC), improving driver goal/impedance estimation (potentially with physiological sensing), and fine-tuning timing and parameters for optimal steering feel and safety.

Interaction type selection is deferred to a higher-level controller and not addressed here. The arbitration rule (simple linear impedance modulation) is a simplification relative to optimal control formulations; realizing literature-defined interaction types would require cost-minimizing control (e.g., nonlinear MPC). The driver goal and impedance estimates are approximate; EKF tuning assumes simultaneous changes in stiffness and damping, limiting capture of asymmetric changes. Model mismatch induces oscillations and steady-state impedance overestimation. Only standard vehicle sensors are used; richer sensing (EEG/EMG) might improve estimation. Inclusion timing from manual intervention to trajectory adaptation requires careful tuning to ensure acceptable steering feel. The driver study sample size (n=5) is small, limiting statistical power. Implementation is bounded by mass-production hardware constraints.