Method of training a machine learning (ML) perception model of an autonomy stack controlling an autonomous vehicle. The autonomy stack has a first processor component for generating a coarse reference trajectory based on detected surrounding obstacles, and a second processor component for refining the trajectory based on perception data, the perception module of the second processor including logic rule-based perception algorithm models and ML perception models. The training method involves identifying objects using rule-based perception model, labelling and annotating the objects, and training the ML perception model. Occupancy grid may be generated, labelling a cell being in occupied, occluded, or free-space state. Occupancy grids of future or previous time points may be temporally paired, such that the grid may include velocity as a state.
Generating a trajectory for an autonomous vehicle
2024-06-26
Patent
Electronic Resource
English