AI-Based Trajectory Inference Powering AFEELA's Autonomous Driving

Sony Honda Mobility engineers Takehiko Hanada and Rintaro Yamaguchi develop trajectory inference for autonomous driving. The technology processes inputs from diverse sensors to understand the surrounding environment, determine vehicle behavior, and create specific motion plans. The approach combines state of the art technology with proprietary modules to address complex traffic scenarios. Automated detection resolves feature interference and supports functional expansion without performance degradation. Simulations automate evaluation of metrics including acceleration and curvature. This process enables quicker identification of bottlenecks and supports generation of high precision trajectories that remain physically consistent. Machine learning operations pipelines automate the sequence from training to evaluation. The changes have reduced overall process time to less than one tenth of the original length. Parallel data operations initiatives have cut the time required to generate training data by more than half. Stored sensor and controller area network data from collection vehicles feed into the latest perception models to produce updated training datasets on a periodic basis. Trajectory inference systems must manage multi functionality across basic and complex driving scenarios while applying state of the art methods. A key risk involves functional interference or degradation, where introduction of new capabilities affects performance of existing ones. Development uses rapid iteration with visualization of degradation patterns, internal model analysis, and automated early detection through simulation. Evaluation covers scenario completion metrics such as target speed and lane deviation along with physical feasibility measures. The training pipeline activates automatically when engineers modify code on GitHub. It then runs training, executes simulations, measures metrics, aggregates results, and presents a final score without manual intervention. Cloud based architecture supports scaling of node capacity as data volume increases. The statements and information are based on development stage data.