In indirect shared control for cooperative driving, the steering commands of the driver and the automation are blended by fixed authority weights on a steer-by-wire platform, and the automation needs the driver's intended reference path to cooperate effectively — but this path is a hidden internal state that the original framework assumed known or identified post hoc under restrictive assumptions. The open problem, posed by Li et al. (IEEE T-ITS 2021) and still unresolved in the 2020–2026 literature, is to infer the driver's intended reference path online from observed steering behavior, integrate that estimate into authority allocation, and arbitrate in a principled way when the driver's and the automation's references disagree, trading off safety, comfort, and driver authority across drivers whose references differ (e.g. curve-cutting versus lane-centering). Follow-up work has produced adaptive authority-allocation schemes and game-theoretic conflict formulations, but none recovers the driver's latent reference path online and closes the loop with the allocation; a 2022 review still lists precise driver-intent recognition as a major open challenge in shared driving.

Background

Shared control of a vehicle means that the human driver and an automated controller simultaneously influence the same steering degree of freedom, as opposed to full manual or full autonomous driving. In haptic shared control the automation applies torque additions to the steering wheel felt by the driver; in indirect shared control, the variant considered here, the physical steering input of the driver and the steering command computed by the automation are combined by a weighted summation before being actuated, in this setting through a steer-by-wire system — a steering architecture with no mechanical linkage between the handwheel and the road wheels, in which wheel angle is set electronically, so that any blend of the two commanded angles can be realized. The weights (λD,λA)(\lambda_D,\lambda_A) of this summation act as control authority: how much of the commanded steering comes from the driver versus the automation.

The source paper (Li, Li, Li, Zhang, Burdet and Cheng, IEEE Transactions on Intelligent Transportation Systems, 2021; preprint arXiv:1704.00866, 2017) developed such an indirect shared controller for lane keeping on fixed-radius curves, validated it in a driving simulator, and modeled the driver with a model predictive control (MPC) scheme — a receding-horizon optimization in which the driver is assumed to choose steering actions minimizing a cost over a short future horizon. A central object in this driver model is the driver's intended reference path: the internal desired lane trajectory the driver is trying to follow, which need not coincide with the lane centerline (some drivers 'cut' curves, tracking a smoother path inside the lane). The experiment showed that the assistance reduced steering effort while preserving lane-keeping performance, but the driver's reference path had to be treated as a known or post-hoc-identified quantity: to make it identifiable from the data, the identification assumed a constant deviation of the driver's reference from the lane center on each fixed-radius curve, and steering power was estimated from steering angles rather than measured directly from torque. Authority weights were fixed and pre-set per trial.

In its conclusions, the paper listed a connected set of unresolved issues. (1) The driver's intended reference path must be inferred online for the automation to cooperate with the driver as actually intended; no practical inference algorithm integrated into the authority allocation or controller design was provided. (2) When the driver's and the automation's references disagree, principled arbitration — when and how to trade off safety, comfort, and driver authority — is unsolved. (3) Drivers differ in their references, and mismatches between the driver's reference and the controller's produced variable steering-energy outcomes in the experiment, so controllers must coordinate or reconcile differing references to consistently reduce driver effort. (4) The identification methodology must be generalized beyond the constant-deviation assumption and validated with direct steering-torque measurement. (5) Systematic methods to choose, adapt, or learn time-varying authority allocations (λD,λA)(\lambda_D,\lambda_A) respecting comfort and performance are needed. These issues form one coherent scientific objective: for indirect shared control to be genuinely cooperative, the automation must know what the driver wants (issues 1, 3, 4) and must know how to act when that differs from what the automation itself would do (issues 2, 5).

Subsequent work in the same line replaced the fixed blending weights with adaptive or learned schemes: a non-zero-sum differential-game formulation of the driver–automation conflict — a differential game is a dynamic game in continuous time in which the players' control inputs jointly steer a shared dynamical system — solved by approximate dynamic programming, a family of numerical methods for dynamic optimization (Li et al., T-ITS 2022); risk-based Gaussian-process dynamic authority allocation on hazardous curves with hierarchical MPC (Zhao et al., T-IV 2023; Gaussian-process regression is a Bayesian nonparametric method that learns a predictive distribution from data); deep-reinforcement-learned human–machine weight allocation trained with the TD3 algorithm (Wang et al., TVT 2024; TD3 is a deterministic policy-gradient actor–critic method for continuous action spaces); and a quadratic-programming allocation — minimization of a quadratic cost under linear constraints — with control-barrier and Lyapunov constraints for smooth authority transitions (Chen et al., T-ITS 2025; a control barrier function is a state function whose invariant level sets enforce safety constraints in a controller, and a Lyapunov function is an energy-like function that decreases along closed-loop trajectories and thereby certifies stability). Adjacent lines treat conflict arbitration via Stackelberg games — a leader–follower game in which one player commits first (Xiong et al., T-ITS 2026; Zhao and Zhu, arXiv 2024, which meta-learns an assistive policy robust to a space of driver utility types), online reinforcement learning with Lyapunov-verified stability (Dang et al., ISA Transactions 2026), and iterative-learning arbitration with driver-in-the-loop results (Yu et al., CCTA 2025). Across all of these, the driver's objective or intention enters as a classified maneuver (e.g. lane-change versus lane-keep), a learned utility parameter, or a behavior-cloned driver agent; the continuous latent intended reference path is not estimated online and closed back into the allocation. A reference-free reinforcement-learning scheme (Yan et al., T-ITS 2024) removes the automation's predefined reference by imitating a learned driver model through behavior cloning (a policy trained by supervised imitation of recorded human steering), but this concerns the machine's reference, not inference of the human's hidden path. Reviews of shared driving (Ansari, Naghdy and Du, T-IV 2022) continue to list precise recognition of driver intention as a major open challenge.

Problem Statement

In the indirect shared-control framework for cooperative driving — where the driver's steering input and the automation's steering command are blended by authority weights (λD,λA)(\lambda_D,\lambda_A) on a steer-by-wire platform, and the driver is modeled as an MPC tracker of an internal intended reference path — solve the cooperation problem left open by Li et al. (2021): (a) infer the driver's intended reference path online from the observed steering behavior and vehicle state, rather than assuming it known or identifying it post hoc under a constant-deviation assumption, and (b) integrate this estimate into the authority allocation so that the automation arbitrates in a principled way — through an explicit, stated trade-off among safety, comfort, and driver authority — between the driver's inferred reference and its own reference when they disagree. The solution should be general across drivers whose intended references differ from the automation's and from each other (e.g. drivers who cut curves versus a lane-centering controller) and should aim at consistently reducing driver steering effort without degrading lane-keeping or safety; it should be validated in a closed loop with driver-model identification that does not presuppose a constant reference deviation. Supporting methodological questions — general identification of the driver's reference path and systematic choice, adaptation, or learning of time-varying authority allocations (λD,λA)(\lambda_D,\lambda_A) respecting comfort and performance — are components of this objective, not separate goals: the arbitration (b) cannot be exercised correctly without the online estimate (a), and the estimate cannot be validated as the driver's true intention without the generalized identification methodology. These sub-results each stand as a partial contribution, but the objective is met only when the online estimate (a) and the allocation with arbitration (b) are demonstrated together in a closed-loop task; simulation or driving-simulator evidence suffices for any claim that does not concern human performance, whereas claims of reduced steering effort or other human-performance outcomes require human-in-the-loop experiments with direct steering-torque-based effort measurement.

References

  1. Renjie Li, Yanan Li, Shengbo Eben Li, Chaofei Zhang, Etienne Burdet, Bo Cheng, "Indirect Shared Control for Cooperative Driving Between Driver and Automation in Steer-by-Wire Vehicles", IEEE Transactions on Intelligent Transportation Systems 22(12):7826-7836, 2021. DOI 10.1109/TITS.2020.3010620. https://doi.org/10.1109/TITS.2020.3010620
  2. Renjie Li, Yanan Li, Shengbo Eben Li, Etienne Burdet, Bo Cheng, "Indirect Shared Control of Highly Automated Vehicles for Cooperative Driving between Driver and Automation", arXiv:1704.00866, 2017. https://arxiv.org/abs/1704.00866
  3. Wenyu Li, Qingkun Li, Shengbo Eben Li, Renjie Li, Yangang Ren, Wenjun Wang, "Indirect Shared Control Through Non-Zero Sum Differential Game for Cooperative Automated Driving", IEEE Transactions on Intelligent Transportation Systems 23(9):15980-15992, 2022. DOI 10.1109/TITS.2022.3146895. https://doi.org/10.1109/TITS.2022.3146895
  4. Xiaobin Zhao, Zhishuai Yin, Zhiwei He, Linzhen Nie, Kang Li, Yuanhao Kuang, Chunyuan Lei, "Indirect Shared Control Strategy for Human-Machine Cooperative Driving on Hazardous Curvy Roads", IEEE Transactions on Intelligent Vehicles 8(3):2257-2270, 2023. DOI 10.1109/TIV.2023.3242762. https://doi.org/10.1109/TIV.2023.3242762
  5. Hongbo Wang, Lizhao Feng, Yuhong Zhang, Juntao Zhou, Haiping Du, "Human-Machine Authority Allocation in Indirect Cooperative Shared Steering Control With TD3 Reinforcement Learning", IEEE Transactions on Vehicular Technology 73(6):7576-7588, 2024. DOI 10.1109/TVT.2024.3352047. https://doi.org/10.1109/TVT.2024.3352047
  6. Yutao Chen, Hongliang Zhang, Haocong Chen, Jie Huang, Bin Wang, Zixiang Xiong, Yuyi Wang, Xiwen Yuan, "Dynamic Control Authority Allocation in Indirect Shared Control for Steering Assistance", IEEE Transactions on Intelligent Transportation Systems 26(3):3458-3470, 2025. DOI 10.1109/TITS.2024.3520107. https://doi.org/10.1109/TITS.2024.3520107
  7. Zhengang Xiong, Bin Li, Lin Wang, Zhenhua Li, Enjian Yao, "Human-Like Decision Making: A Human–Machine Shared Steering Control System With Adaptive Authority Allocation", IEEE Transactions on Intelligent Transportation Systems 27(2):2268-2280, 2026. DOI 10.1109/TITS.2025.3620893. https://doi.org/10.1109/TITS.2025.3620893
  8. Viet Thang Dang, Duy Lam Mac, Duc Thinh Le, Tung Lam Nguyen, Van Trong Dang, "An online reinforcement learning-based conflict resolution approach for human–vehicle cooperative systems", ISA Transactions 169:391-407, 2026. DOI 10.1016/j.isatra.2025.12.017. https://doi.org/10.1016/j.isatra.2025.12.017
  9. Zihan Yu, Aleksandra Dudek, Patrick Linford, Ian Holbrook, Scott James, Matthew Castanier, Kira Barton, Chris Vermillion, "Iterative Learning-Based Arbitration in Shared Vehicle Control: Methodological Formulation and Driver-in-the-Loop Results", 2025 IEEE Conference on Control Technology and Applications (CCTA), pp. 1-7, 2025. DOI 10.1109/CCTA53793.2025.11151407. https://doi.org/10.1109/CCTA53793.2025.11151407
  10. Shahzeb Ansari, Fazel Naghdy, Haiping Du, "Human-Machine Shared Driving: Challenges and Future Directions", IEEE Transactions on Intelligent Vehicles 7(3):499-519, 2022. DOI 10.1109/TIV.2022.3154426. https://doi.org/10.1109/TIV.2022.3154426
  11. Yuhan Zhao, Quanyan Zhu, "Stackelberg Meta-Learning Based Shared Control for Assistive Driving", arXiv:2403.10736, 2024. https://arxiv.org/abs/2403.10736
  12. Liang Yan, Xiaodong Wu, Chongfeng Wei, Sheng Zhao, "Human-Vehicle Shared Steering Control for Obstacle Avoidance: A Reference-Free Approach With Reinforcement Learning", IEEE Transactions on Intelligent Transportation Systems 25(11):17888-17901, 2024. DOI 10.1109/TITS.2024.3420894. https://doi.org/10.1109/TITS.2024.3420894
  13. Mingjun Li, Haotian Cao, Xiaolin Song, Yanjun Huang, Jianqiang Wang, Zhi Huang, "Shared Control Driver Assistance System Based on Driving Intention and Situation Assessment", IEEE Transactions on Industrial Informatics 14(11):4982-4994, 2018. DOI 10.1109/TII.2018.2865105. https://doi.org/10.1109/TII.2018.2865105
  14. Chen Zhou, Fan Zhang, Edric John Cruz Nacpil, Zheng Wang, Fei-Xiang Xu, "Driver Steering Intention Prediction for Human-Machine Shared Systems of Intelligent Vehicles Based on CNN-GRU Network", Sensors 25(10):3224, 2025. DOI 10.3390/s25103224. https://doi.org/10.3390/s25103224