Online Maneuver Learning and Its Real-Time Application to Automated Driving System for Obstacles Avoidance

Tatehara, Takumi; Nagahama, Akihito; Wada, Takahiro · 2023 · Crossref

DOI: 10.1109/tiv.2022.3146622

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of enhancing driver comfort and trust in Automated Driving Systems (ADS) by enabling real-time, online learning of individual driver preferences. While existing methods rely on offline learning from previously collected data, they cannot immediately adapt to a driver’s specific maneuvers during a driving event. The authors propose an on-demand online learning method focused on obstacle avoidance, allowing the ADS to learn preferred paths during the maneuver and apply the results in real time. This approach aims to facilitate smooth transitions between automated and manual control, thereby improving the system's adaptability and user acceptance. The proposed method utilizes B-spline interpolation for path planning due to its low computational cost, enabling updates in less than 0.01 seconds. The system employs a haptic shared control mechanism where steering authority is dynamically transferred between the driver and the ADS based on a "cooperative state" derived from pseudo-workload calculations. When a driver intervenes (State II), the ADS reduces its control gain and learns the traveled path by adjusting three specific control points of the B-spline curve: the start of avoidance, the lateral gap beside the obstacle, and the return to the lane center. These parameters are updated in real time using proportional-derivative-like control laws. If the driver ceases intervention, indicating satisfaction with the adjusted path (State III), authority returns to the ADS. The study was evaluated using a driving simulator with ten participants. The experimental scenario involved an ego vehicle avoiding a parked obstacle at speeds of 40 and 60 km/h. The design compared conditions before learning (no intervention, intervention without learning, and intervention with online learning) and after learning (observing the learned path, intervening without further learning, and iterative learning over three trials). Performance was measured objectively via driver steering torque and subjectively through questionnaire surveys assessing comfort and trust. Results demonstrated that the proposed ADS significantly improved driver comfort and trust compared to non-learning systems. The system successfully learned individual driver preferences during avoidance maneuvers, as evidenced by reduced driver torque in subsequent trials and the convergence of B-spline control points to values reflecting driver intent. Specifically, the real-time adjustment of the path allowed drivers to recognize the system's adaptation, leading to smoother authority transfers. The findings confirm the feasibility of online on-demand learning for ADS, suggesting that such methods can enhance convenience and facilitate fast adaptation to diverse traffic situations and driver styles.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• automation surprise
• in vehicle coaching
• steering pattern

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model