Nova : Open-Source Self-Driving Software Stack

Problem:

Despite the numerous players and success in driverless technologies, there remains no good platform for academic researchers who want to develop technologies for this application but have no way of testing them in a complete system. The competitive and fast-paced nature of the race towards self-driving vehicles has effectively kept most academic researchers at arms-length from the cutting edge. Autonomous Vehicle(AV) software stacks are proprietary and the few projects that aim to be open-source are either still in development or aim to bring a commercially competitive solution to market. What is lacking is a robust and easy-to-adopt/maintain /run AV stack that researchers can employ to test new technologies that might advance what is possible for self-driving vehicles.

Solution:

Nova is the applied self-driving arm of Dr. Ruth's research in the Applied Systems Lab. The Nova project is composed of an entirely undergraduate team to develop and refine our in-house ROS(Robot Operating System) -based AV software stack. Our ongoing development targets building out core driverless algorithms in a simple and modular way such that reseachers who use our platform can swap out, rewrite, or add new components to evaluate their own, for example, sensing technologies, algorithms, or hardware.

 

Technology Readiness:

Level 6

Technologies:

Video/LIDAR/IMU/GNSS sensing, machine learning perception algorithms, behavior planning, control

 

Finding Vulnerabilities in Autonomous Arial Vehicles

Problem:

Autonomous Arial Vehicles (AAVs), i.e., drones, have been increasingly adopted in commercial and industrial applications, such as package delivery and surveillance. At the same time, there has been an increasing rate of AAV accidents with critical consequences, such as vehicle malfunction and crashes. Hackers have also demonstrated numerous attacks exploiting security vulnerabilities in AAVs to crash and take control of the vehicles maliciously. At the core of these incidents is the lack of a systematic approach to testing and analyzing AAVs. Due to a semantic gap between the cyber and physical domains of the systems, it is challenging to correlate the causes of vehicle accidents/attacks in the cyber domain (e.g., a software bug) with the resulting symptoms in the physical domain (e.g., a vehicle crash), which are critical for testing and analyzing AAVs accurately and retrofitting them to avoid the same incidents.

Solution:

RetroV offers automated testing and analysis tools to find bugs and vulnerabilities in AAVs and retrofit their design to prevent accidents and possible attacks. Our tools discover hidden physical misbehaviors of AVs under test and identify the root cause with near-zero human intervention. We achieve this by leveraging a high-fidelity drone simulator to generate a virtually infinite set of test cases as well as real drones for real-world analysis. During testing, our tool automatically monitors the vehicle’s internal states in real time and records them to trace the root cause in case a misbehavior is detected. When an incident occurs, our tool effectively closes the semantic gap between the misbehavior (in the physical domain) and the responsible system code (in the cyber domain) by analyzing the causal relationships across vehicle states and mapping them to the code. Our tools have successfully discovered over 90 new vulnerabilities in two AVs. These systems power many popular commodity drones in the real world.

Technology Readiness:

Level 6

Technologies:

Drones, Software Testing, Program Analysis, Cyber-Physical Systems

Robust and Risk-Aware Autonomous Vehicles

Problem:

Safety is a critical issue for robotic and autonomous systems that must traverse through uncertain environments. More sophisticated motion planning and control algorithms are needed as environments become increasingly dynamic and uncertain to ensure safe and effective autonomous behavior. Safely deploying robots in such dynamic environments requires a systematic accounting of various risks both within and across layers in an autonomy stack from perception to motion planning and control. Many widely used motion planning algorithms have been developed in deterministic settings. However, since motion planning algorithms must be coupled with the outputs of inherently uncertain perception systems, there is a crucial need for more tightly coupled perception and planning frameworks that explicitly incorporate perception and prediction uncertainties.

Solution:

We have developed a distributionally robust incremental sampling-based motion planning framework that explicitly and coherently incorporates perception and prediction uncertainties. We utilize advanced techniques in robust and stochastic optimization to ensure that safety constraints are satisfied despite these inherent uncertainties. We have demonstrated our proposed algorithms and approaches using nonlinear vehicle models in an open urban driving simulator with static and dynamic obstacles} and show that risk bounded motion planning can be achieved effectively for nonlinear robotic systems. We are currently exploring the development of full autonomy stacks that tightly integrate layers from perception and prediction to planning and control and incorporate modern artificial intelligence and machine learning technologies.

Technology Readiness:

3-4 Level

Technologies:

Stochastic and robust optimization; motion planning under uncertainty; vehicle dynamics; feedback control;. artificial intelligence; machine learning.