CommonRoad-to-Autoware 2.0: Simplifying Sim-to-Real Transfer of Behavior Planning in Autonomous Driving

Implementing, orchestrating, and validating behav- ior planning algorithms on real vehicles is crucial for real- world autonomous driving. While open-source stacks, such as Autoware, enable real-world deployment, their complexity makes integrating new methods difficult. As a result, many novel approaches are evaluated only in simulation, for example using the CommonRoad benchmark suite. Our previous work CR2AW – a publicly available interface between the CommonRoad framework and Autoware – already significantly simplifies the sim-to-real transfer of motion planning research. In this paper, we present CR2AW-2.01, which extends the capabilities of our previous work to the behavior planning layer, enabling a holistic sim-to-real transfer across the entire planning stack. Motivated by practical concerns, our architecture combines a finite state machine for managing components of the planning stack with a behavior tree for structured low-level decision-making. We showcase CR2AW-2.0 by implementing four behaviors: reacting to traffic lights, adjusting the velocity in narrow situations, lane keeping, and decelerating into standstill during a system failure. Our experiments both in simulation and on our research vehicle showcase the usefulness of CR2AW-2.0.     «

Implementing, orchestrating, and validating behav- ior planning algorithms on real vehicles is crucial for real- world autonomous driving. While open-source stacks, such as Autoware, enable real-world deployment, their complexity makes integrating new methods difficult. As a result, many novel approaches are evaluated only in simulation, for example using the CommonRoad benchmark suite. Our previous work CR2AW – a publicly available interface between the CommonRoad framework and Autoware...     »