Cooperative Autonomous Driving
Team AnnieWAY is formed by researchers working on automated driving at Karlsruhe Institute of Technology and FZI Research Center for Information Technology, in Germany. Since its foundation in 2001, the research group has achieved many milestones in automated driving, such as DARPA Urban Challenge 2007, Grand Cooperative Driving Challenge 2011, the fully autonomous completion of the Bertha Benz Memorial Route in 2013, and Grand Cooperative Driving Challenge 2016.
Our automated vehicle Bertha (black car) in platoon communicating with other vehicles via v2v-communications.
Cooperative intersection crossing: Bertha (black car) accelerates after it is sure that the grey car will leave the intersection area.
We participated in the Grand Cooperative Driving Challenge 2016 and took the second place in the overall score. Our optimization-based motion planner played a dominant role in getting the excellent final score. Our virtual validation framework enabled us for intensive testing before the challenge and to tune the relevant configurations. More detailed information can be found on our paper.
Motion Planning and Decision-Making
Ensuring safe interactions and planning collision-free trajectories are essential for the widespread adoption of autonomous vehicles. The uncertainties in this process stem from various sources, including limited sensor range, noisy sensor data, and objects occluded in the environment. Additionally, autonomous vehicles must adeptly handle the unpredictable behaviors of other traffic participants, including those who may disregard traffic rules. To maintain safe navigation in complex and dynamic scenarios, these uncertainties must be effectively addressed. Managing these uncertainties is essential to achieve interactive and reliable behavior.
Surface-of-no-return
Safety inspected in speed–time–path (v − t − s) space. Any motion below the red-depicted surface-of-no-return can come to a full stop before the given intersection. The black line represents an unsafe motion as it leaves the red surface at the green point, at the point-of-no-return.
Perception and prediction uncertainty
An automated vehicle should be able to compensate rule-noncompliant behavior of other traffic participants when approaching to an intersection where it has the right-of-way. While being proactive, it should not move too defensive.
Limited visibility
For various sensor ranges (30 m − 110 m) the speed profile along the path of an intersection scenario. The vertical dashed red line corresponds to the intersection point.
Passive information gathering
Motion planning involves decision-making among alternative maneuvers. The choice among maneuvers depends on the uncertainties and associated risks of the maneuver variants. In cases where the uncertainty is high, many methods resort to defensive motion. However, if the feasibility of a contingecy maneuver is guaranteed, a decision can be postponed to a later time with the expectation that more information will be available in the future. We call this type of motion maneuver-neutral and the action as passive information gathering.
The next figure shows a use case highlighting the benefit of planning neutral trajectories in case of a phantom object detection. The depicted driving scenario on the right is mapped on the path-time diagram on the left. The planner considers alternative maneuvers (A and B) in a combined fashion and plans a motion (C) that is formed by the probabilities of those alternatives. This approach enables a smooth reaction to low-probability events while ensuring collision avoidance if any of these events actually occur.
Closed-Loop Simulation Framework
Probabilistic Prediction and Planning for Intelligent Vehicles (P3IV) Simulator is a simulation framework I have developed for motion prediction and planning of autonomous vehicles.
Focus:
- Algorithm development for prediction and planning
- Consideration of uncertainties and limited visibility
- Covers multi-agent interactions
- Provides utility libraries for prediction and planning
- Allows both open-loop and closed-loop simulation
Key Features:
- Catkin package structure: seamless integration into ROS
- Implemented in Python and C++; wrapped with PyBind
- Build on the HD map library Lanelet2
- Bindings to simulate with real-world drone datasets
- BSD 3-Clause license
Motion Control
Planned trajectories typically provide a coarse-grained reference for vehicle navigation. These references must be accurately tracked and stabilized against unmodeled dynamics and external disturbances. In collaboration with colleagues at MAN Truck & Bus SE, we are working on addressing these challenges. Our work includes topics around bi-level stabilization, sparsification, and stochastic MPC.
Self-Awareness
Ensuring safety is crucial for the widespread adoption of autonomous vehicles. While state-of-the-art autonomous vehicles perform well under ideal conditions, their operation becomes significantly challenging in severe weather or when other road users fail to follow traffic rules. The EU Horizon 2020 RobustSENSE project aims to address these challenges. This project focuses on creating a holistic processing architecture to enhance the operational robustness and reliability of autonomous vehicles. My role in this project involved developing a new concept and design for the system architecture.
The Performance Assessment units employ metrics reflecting quality of the obtained results. The System Performance Assessment module continuously evaluates the performance status of the whole vehicle to be able to react by either function degradation or component adaptation.
The multi-server ROS system acts as one integrated system for peripheral devices like sensors and actuators. The modular and extendible structure of ROS allows for adapting the system to short-term modifications that makes it especially powerful for prototyping purposes.