Driver safety is aided by multiple driver assist systems in modern cars, such as anti-lock braking and lane departure warning systems. The prevention or amelioration of a certain class of accidents is dependent on rapid situation assessment and resultant highly responsive evasive maneuver. Increasingly capable sensors are improving situation awareness. However, current driving planners/controllers for autonomous cars fail to generate a safe plan, particularly in emergency situations. The main reason is that such planners are designed to produce comfortable trajectories for human passengers under normal driving conditions. This project aims to detect conditions under which potentially high-performance evasive maneuver is necessary to avoid collision, and to invoke a special planner/controller evasive maneuver mode in order to react safely to such conditions.
Problem statement Driver safety is aided by multiple driver assist systems in modern cars, such as anti-lock braking and lane departure warning systems. The prevention or amelioration of a certain class of accidents is dependent on rapid situation assessment and resultant highly responsive evasive maneuver. Increasingly capable sensors are improving situation awareness. However, current driving planners/controllers for autonomous cars fail to generate a safe plan, particularly in emergency situations. The main reason is that such planners are designed to produce comfortable trajectories for human passengers under normal driving conditions. This project aims to detect conditions under which potentially high-performance evasive maneuver is necessary to avoid collision, and to invoke a special planner/controller evasive maneuver mode in order to react safely to such conditions. Application The developed methods would apply and could be invoked whenever emergency conditions are detected. In manual driving, this could occur if the driver’s attention is wandering, or his limited situation awareness is not sufficient to take in the entirety of a dangerous situation. An automatic evasive maneuver capability in that case would be like an extreme extension of the anti-lock braking already present on production vehicles, which does not require explicit user activation. In fully autonomous driving, the evasive maneuver capability would invoke extreme performance capabilities of the vehicle and its planner/controller not active under normal conditions. Approach The proposed work has two components: detection and planning/execution. The detection component will be dealt with at a simple level in the proposed work, though it is a complex topic in its own right and should be elaborated in a separated and related project. In the proposed work, detection will be based on the combination of road topology, prediction of the future course of all relevant vehicles (probabilistic where multiple road topology options exist), and knowledge of human response times. Issues such as focusing of sensor attention and driver attention/gaze monitoring should be considered in separate work, but we will neglect them. Given our simplifying assumptions, we will develop techniques for the avoidance of false positives in determining the necessity for evasive maneuver. The main work will be in the planning/execution area. The typical autonomous driving framework is hierarchical, with the following general levels: high-level reasoning, motion planning, tracking control, and actuation control. This framework can introduce significant delays in execution which are tolerable in normal driving, but not in high-performance evasive maneuver. We will perform more sophisticated planning-control integration using a model-based planning/control scheme by performing the following tasks: 1. Abstract away "tracking control" and let the output of the motion planning directly control the actuation, removing the lag involved in an extra tracking control layer. 2. To achieve that, account in the planner for a more realistic system model (actuation model + vehicle model). 3. Possibly add GPU acceleration in order to allow the planner to consider more trajectory options in real time. Validation Initial validation will be performed by devising a small set of canonical scenarios that reflect the most frequent accident circumstances, which occur at intersections. These will be exhaustively investigated in simulation. They will then be validated in indoor tests on a small-scale real-world vehicle, a RC car. The car’s dynamics will be modeled and included in the planning/control loop. Realistic testing of autonomous evasive maneuver on an actual vehicle is challenging for multiple reasons: 1) a vehicle retrofitted for autonomy is needed; 2) fully authentic testing requires the creation of dangerous situations; 3) such testing has to be conducted at least initially on a controlled, closed course.We have access to an autonomous vehicle in the form of CMU’s autonomous Cadillac SRX. We will attempt to devise low-speed scenarios allowing validation of the basic approach. It may be possible to scale the results for partial validation, but high-speed vehicle dynamics will not come into play for safety reasons and will only be able to be validated in highly controlled test situations (see Deployment Plan below).
Duration: 1 year (Jan. 1 –Dec. 31, 2015)High-level schedule:Mar. 31: Architecture and theoretical developmentJune 30: Validation in softwareSept. 30: Validation on indoor hardware (RC car)Nov. 31: Validation on SRXDec. 31: Final report, documentation of hardware/software
Even the deployment of something as relatively innocuous as anti-lock braking systems faced some opposition in the form of user discomfortat not being in full control of the vehicle. This applies much more strongly to a capability as invasive of a driver’s control prerogatives as autonomous evasive maneuver. The best path for deployment is through our research sponsor and partner GM. They have the controlled testing facilities and procedures that allow rigorous and maximally safe testing of such a capability. The deployment plan is therefore to: 1. Validate a basic capability to perform autonomous evasive maneuver during the course of the proposed project. 2. Begin to transition this capability to GM R&D through the technology transfer mechanism already in place via the GM-CMU Autonomous Driving Collaborative Research Lab (AD-CRL). 3. The technology can then be matured through GM’s proven processes to the point that it may become suitable for deployment on passenger vehicles.
Accomplishments - Creation of a novel high-performance evasive maneuver planner - Validation of the planner in simulation, on a small-scale vehicle, and partially on a full-scale vehicle - Documentation of the results in a form appropriate for hand-off to a partner capable of full-scale deployment Metrics - Ability to deal with 90% of the most frequent avoidable accident scenarios - Reduction of total planning latency to a value at least comparable to, if not smaller than, human reaction times (subsequent work should reduce it further)
Name | Affiliation | Role | Position | |
---|---|---|---|---|
jmd@cs.cmu.edu | Dolan, John | Robotics Institute | PI | Faculty - Research/Systems |
tianyu@cmu.edu | Gu, Tianyu | ECE | Other | Student - Masters |
Type | Name | Uploaded |
---|---|---|
Final Report | 2015_Dolan_-_UTC_Final_Report.pdf | May 11, 2018, 4:19 a.m. |
Publication | Human-like planning of swerve maneuvers for autonomous vehicles | April 19, 2021, 6:11 a.m. |
Publication | Automated tactical maneuver discovery, reasoning and trajectory planning for autonomous driving | April 19, 2021, 6:12 a.m. |
Publication | Safe Planning for Self-Driving Via Adaptive Constrained ILQR | April 19, 2021, 6:14 a.m. |
Publication | Safe planning and control under uncertainty for self-driving | April 19, 2021, 6:15 a.m. |
No match sources!
No partners!