Yuseung Na

Context: End-to-end autonomous driving (E2EAD) jointly trains the core sensor-to-motion mapping, so requirements are not exposed through module-like responsibility boundaries and module contracts cannot serve as the sole unit for requirements allocation.Objective: This study identifies the requirement-related concerns that E2EAD development must specify and designs a requirements engineering process connecting requirements derivation to verification and assurance evidence.Method: A framework-informed evidence-mapping study of 83 publications, using LLM-assisted discovery and human screening, extraction, and coding, produced 17 requirement element groups and informed five process design requirements. RESHAPE combines this structure with a System–Model–Data Layer (SMDL) V-Process that derives requirements top-down, promotes controlled baselines bottom-up through evidence gates, and makes proxy adequacy and requirement-to-version traceability explicit. Evaluation combined retrospective analyses of a joint industrial program, requirements authoring for a non-motivating SAE Level 2 highway feature, and an assessment by six reviewers independent of the framework's design.Results: The element structure represents all ten documented problems and, across them, all 17 groups; seven problems require cross-layer diagnosis, for which typed links and a worked example define an analytical path. Inter-coder agreement for the retrospective element coding was κ=0.81. The highway-feature application produced 43 requirements and instantiated the cross-layer derivation links. Four practitioners and two external academics supported the framework's face validity within this domain while identifying trace-link maintenance as the main implementation constraint.Conclusion: Treating System, Model, and Data as separately specified, evidence-gated, and version-traceable layers provides a basis for specifying E2EAD requirements and a defined diagnostic structure for cross-layer failures. Establishing effectiveness requires prospective, multi-organization studies.

Learning-based motion planners infer driving behavior from data, reducing reliance on hand-crafted rules. However, closed-loop deployment remains difficult because their outputs can fluctuate across planning cycles, violate vehicle dynamics and friction limits, and degrade ride comfort through unnecessary acceleration and steering oscillations. Prior safety filters and low-level control corrections can mitigate immediate violations, but they do not provide a unified mechanism that enforces trajectory-level consistency, executability, and risk-aware validation at the planner output. This paper proposes a refinement-and-verification trajectory processing pipeline that improves closed-loop execution without replacing the underlying planner. The pipeline combines a Kalman filter–based consistency refinement across planning cycles, an NMPC-based Dynamic–Safety–Comfort refinement that enforces vehicle dynamics together with safety and comfort constraints, and a risk-aware verification and selection stage based on friction limits, collision risk, and driving-quality metrics. On the NuPlan closed-loop benchmark, the full pipeline improves ride-comfort metrics by more than 10% and increases collision- and TTC-related safety metrics by about 4%, yielding a consistent gain in the overall planning score. Additional evaluation in CarMaker and real-vehicle experiments shows that the refined trajectories remain dynamically feasible and can be stably tracked without sacrificing task-level driving performance. These results support the role of a structured trajectory-level intermediate layer for bridging learning-based planning and real-world execution. Supplementary real-vehicle comfort demonstration videos are available at https://github.com/ailab-hanyang/ML-Planning-Refinement-Verification.

Recent advances in learning-based planners have accelerated progress toward automated driving. However, safe deployment still depends on handling rare safety-critical situations, to which such planners remain vulnerable because these events are underrepresented in naturalistic driving data. Existing adversarial scenario generation methods address this gap but suffer from three limitations: they produce either inevitable collisions or insufficiently threatening situations rather than useful near-miss scenarios; they can generate trajectories that violate physical motion constraints; and their generated datasets are rarely validated across diverse learning-based planners. We propose PhysCrit, a three-stage framework that augments nuPlan driving logs with physically feasible safety-critical scenarios. First, it classifies the space ahead of the ego vehicle into inevitable-collision, emergency-braking, and safe regions based on the required deceleration, and selects an attack target only from the emergency-braking region. Second, it solves two sequential model predictive control (MPC) problems to generate an adversarial agent trajectory and an ego evasive trajectory, both constrained by a kinematic bicycle model with coupled acceleration, jerk, and steering-rate limits. Third, it applies a three-step validation retaining only scenarios with an effective attack and a physically feasible evasion. We evaluate PhysCrit on two state-of-the-art, architecturally distinct planners, PLUTO and Diffusion Planner. PhysCrit-augmented training improves safety in safety-critical scenarios without substantially degrading general-driving performance on the nuPlan Test14-Hard and Test14-Random benchmarks. These results indicate that physically feasible near-miss generation offers a scalable and reproducible basis for safety-critical training and validation of automated driving systems.

An electronically controlled suspension with road preview enhances ride comfort and driving stability. For detecting road irregularities such as speed bump, look-ahead sensors such as camera and Light Detection And Ranging (LiDAR) sensors have been used. However, these preview systems perform poorly in night and adverse weather conditions. To overcome these limitations, this paper proposes a novel machine learning-based speed bump prediction system that predicts remaining distance to and height of speed bumps using in-vehicle data, which is robust to external environment and represent driving pattern of driver. The proposed system consists of three components: machine learningbased bump predictor, automatic training dataset generator, and cloud-based continuous learning system. The effectiveness of the proposed system was confirmed through tests with varied datasets in different drivers and locations. The prediction of the distance to the speed bump demonstrated a Root Mean Square Error (RMSE) of 6.37 meters, while the height prediction showed an RMSE of 0.9 cm. Additionally, after applying cloud-based continuous learning for driver-specific personalization, the RMSE for the distance prediction improved to 5.96 meters. Furthermore, experiments conducted under challenging conditions showed that our system outperformed existing systems.

A computing system for autonomous vehicles must efficiently process vast amounts of data from various sensors in real-time and have a backup design to handle system failures. Adding more computing devices improves performance and redundancy, but increases costs and energy consumption. With the development of vehicle-to-everything (V2X) communication technologies and edge computing, such as multiaccess edge computing (MEC), computation offloading has been introduced. This process involves the use of cloud or edge servers to handle calculations normally performed by on-board computers, thereby reducing their workload and offering redundancy. This article proposes an offloading strategy that optimizes the allocation of software components (SWCs) in autonomous driving software. By optimizing SWC allocation, SWCs can be effectively assigned to specific computing units. This article established a cost function and constraints based on SWC-specific features to address the offloading decision problem as an allocation optimization issue. This article also defines safety-related metrics, including the response time requirements and failure risks of SWCs, to set criteria for offloading decisions. The proposed SWC optimization method is tested and validated using a real autonomous driving vehicle demonstration involving both cloud and vehicular edge environments.

In the realm of autonomous driving, motion planning for the ego vehicle necessitates the prediction of surrounding vehicles’ motions. This prediction traditionally relies on object tracking modules containing several sensors to gauge vehicle positions and velocities. However, existing physical or maneuver-based model approaches overlook important aspects of vehicle interactions that significantly affect actual vehicle movement. Ignoring these interactions can lead to inadequate ego vehicle’s motion planning. Addressing this gap, this paper proposes a novel approach: the Collision Probability Field (CPF)-based interaction-aware longitudinal motion prediction. Our methodology uniquely integrates the CPF, derived from the uncertainty of sensing information, to account for the probabilistic state of vehicle positions and velocities. This allows the prediction algorithm to consider not just the static data, but also the dynamic interactions between vehicles such as collision. Our approach was tested in various scenarios, including lane changes with an approaching vehicle from behind and different driver behavior models in real-world conditions. Our findings demonstrate a significant improvement in prediction accuracy for the motion planning of ego vehicle, highlighting the importance of interaction-aware predictions in autonomous driving systems.

As an alternative to traditional remote controller, research on vision-based hand gesture recognition is being actively conducted in the field of interaction between human and unmanned aerial vehicle (UAV). However, vision-based gesture system has a challenging problem in recognizing the motion of dynamic gesture because it is difficult to estimate the pose of multi-dimensional hand gestures in 2D images. This leads to complex algorithms, including tracking in addition to detection, to recognize dynamic gestures, but they are not suitable for human–UAV interaction (HUI) systems that require safe design with high real-time performance. Therefore, in this paper, we propose a hybrid hand gesture system that combines an inertial measurement unit (IMU)-based motion capture system and a vision-based gesture system to increase real-time performance. First, IMU-based commands and vision-based commands are divided according to whether drone operation commands are continuously input. Second, IMU-based control commands are intuitively mapped to allow the UAV to move in the same direction by utilizing estimated orientation sensed by a thumb-mounted micro-IMU, and vision-based control commands are mapped with hand’s appearance through real-time object detection. The proposed system is verified in a simulation environment through efficiency evaluation with dynamic gestures of the existing vision-based system in addition to usability comparison with traditional joystick controller conducted for applicants with no experience in manipulation. As a result, it proves that it is a safer and more intuitive HUI design with a 0.089 ms processing speed and average lap time that takes about 19 s less than the joystick controller. In other words, it shows that it is viable as an alternative to existing HUI.

Car-following control is a fundamental application of autonomous driving. This control has multiple objectives, including tracking a safe distance to a preceding vehicle and enhancing driving comfort. Model Predictive Control (MPC) is a powerful method due to its intuitiveness and capability to cover multiple objectives. MPC determines the relative importance of objectives through a set of weight factors, depending on which, the controller’s behavior changes even if the traffic situations are the same. However, determining the optimal weight is not a trivial problem because there is no benchmark to evaluate the performance of the weight, and searching for weight factors with repeated driving experiments is time-consuming. To solve this problem, we proposed an automatic tuning method to determine the weights of the MPC based on personal driving data. Personal driving data under naturalistic driving conditions provide car-following situations and driver’s behaviors. These data can generate a reference model to represent the driver’s driving style. Based on this model, the proposed method defined the automatic tuning problem as an optimization problem that minimizes the difference between the reference and the controller’s response using the optimal weight factors. This optimization problem was solved using the Particle Swarm Optimization algorithm. The proposed method was implemented with an embedded optimization coder in an offline fashion. Its performance was evaluated using personal driving data. From this, the proposed method can reduce the effort and time required for an engineer to find the optimal weight factors.