Belief State Planning for Autonomous Driving: Planning with Interaction, Uncertain Prediction and Uncertain Perception

This thesis presents a behavior planning algorithm for automated driving in urban environments with an uncertain and dynamic nature. The uncertainty in the environment arises by the fact that the intentions as well as the future trajectories of the surrounding drivers cannot be measured directly but can only be estimated in a probabilistic fashion. Even the perception of objects is uncertain due to sensor noise or possible occlusions. When driving in such environments, the autonomous car must predict the behavior of the other drivers and plan safe, comfortable and legal trajectories. Planning such trajectories requires robust decision making when several high-level options are available for the autonomous car.
Current planning algorithms for automated driving split the problem into different subproblems, ranging from discrete, high-level decision making to prediction and continuous trajectory planning. This separation of one problem into several subproblems, combined with rule-based decision making, leads to sub-optimal behavior.
This thesis presents a global, closed-loop formulation for the motion planning problem which intertwines action selection and corresponding prediction of the other agents in one optimization problem. ... mehrThe global formulation allows the planning algorithm to make the decision for certain high-level options implicitly. Furthermore, the closed-loop manner of the algorithm optimizes the solution for various, future scenarios concerning the future behavior of the other agents. Formulating prediction and planning as an intertwined problem allows for modeling interaction, i.e. the future reaction of the other drivers to the behavior of the autonomous car.
The problem is modeled as a partially observable Markov decision process (POMDP) with a discrete action and a continuous state and observation space. The solution to the POMDP is a policy over belief states, which contains different reactive plans for possible future scenarios. Surrounding drivers are modeled with interactive, probabilistic agent models to account for their prediction uncertainty. The field of view of the autonomous car is simulated ahead over the whole planning horizon during the optimization of the policy. Simulating the possible, corresponding, future observations allows the algorithm to select actions that actively reduce the uncertainty of the world state. Depending on the scenario, the behavior of the autonomous car is optimized in (combined lateral and) longitudinal direction. The algorithm is formulated in a generic way and solved online, which allows for applying the algorithm on various road layouts and scenarios.
While such a generic problem formulation is intractable to solve exactly, this thesis demonstrates how a sufficiently good approximation to the optimal policy can be found online. The problem is solved by combining state of the art Monte Carlo tree search algorithms with near-optimal, domain specific roll-outs.
The algorithm is evaluated in scenarios such as the crossing of intersections under unknown intentions of other crossing vehicles, interactive lane changes in narrow gaps and decision making at intersections with large occluded areas. It is shown that the behavior of the closed-loop planner is less conservative than comparable open-loop planners. More precisely, it is even demonstrated that the policy enables the autonomous car to drive in a similar way as an omniscient planner with full knowledge of the scene. It is also demonstrated how the autonomous car executes actions to actively gather more information about the surrounding and to reduce the uncertainty of its belief state.