Motion Planning for Autonomous Driving

Trajectory planning is an essential task of autonomous vehicles. Map- and sensor-based data form the basis to generate a trajectory that serves as a target value to be tracked by a controller. Besides comfort aspects, the feasibility and possible collisions must be taken into account when generating the target trajectory.

An important aspect of trajectory and in general motion planning is predicting the behavior of the surrounding road users. The majority of trajectory planning algorithms are based on the assumption that other road users behave independently of the generated trajectory. The predicted trajectories are assumed to be fix during the planning process. Very conservative and in some situations even dangerous behaviors can occur due to the ignorance of mutual dependencies. This motivates interactive trajectory planning, where the reaction of other road users is taken into account already during the planning phase. Prediction and planning should therefore no longer be treated as a sequential process, but rather be combined. The difficulty is to consider the reaction of the surrounding road users for a long planning time horizon.

One way of representing the behavior of human drivers is to use probabilistic models. Data-driven methods help to generate probability distributions of possible behaviors. To take interactions into account, such models can then be coupled. Other approaches are based on game theory and model the interactions as dynamic games. Current challenges here are the computational effort and the consideration of uncertainties. Furthermore, learning approaches that can adapt to the variability of road users are investigated. Possible methods are reinforcement learning and classification algorithms.