Presentation
Achieving Verifiable Autonomy: Is Design Automation the Golden Key?
DescriptionIn many domains, and in particular in the automotive domain, we have already overcome several technical challenges on the path towards implementing autonomous systems. However, practical realization and mass adoption of autonomous systems and vehicles still remains elusive. One of the fundamental remaining obstacles is to ensure that the system behaves correctly in all possible scenarios – a goal that needs to be achieved via formal or semi-formal verification.
But the complexity of most autonomous systems makes this a very hard final challenge. In particular, traditional abstractions and well-tested principles that follow the "separation of concerns” paradigm to manage complexity, no longer seem to work. For example, fully verifying machine learning (ML) based perception processing subsystems in isolation might not be feasible. But when such perception processing is viewed in conjunction with the control components they feed into, full verification of the ML components might not be necessary to provide "global” correctness guarantees. Similarly, while high-level models of control components can be formally verified today, verifying software implementations synthesized from such models, especially when they are running on complex distributed architectures is more difficult. Building simulations and virtual prototypes are both difficult and expensive and still do not offer the kind of guarantees that are necessary. Along the same lines, while the real-time systems literature provides an array of formal methods for timing analysis and scheduling, few of them scale to the complexity found in real-life autonomous systems.
In summary, addressing the verification challenge for autonomous systems requires more holistic approaches that can "connect” different subsystems together for design, optimization, and verification. Such cross-layer verification requires design automation and suitable design tools. Do such tools and methods already exist? Which gaps in the availability of design automation tools are already recognized in the industry? What kinds of scientific challenges in developing such necessary tool support is recognized by the academic research community? And most fundamentally, does design automation indeed hold the key to achieving verifiable autonomy without overprovisioning resources?
The goal of this panel would be to discuss and debate on these questions. It will shed light on perspectives of the academic research community and also that of the industry. The outcome of this panel discussion would be an actionable research agenda for embedded systems and design automation researchers.
But the complexity of most autonomous systems makes this a very hard final challenge. In particular, traditional abstractions and well-tested principles that follow the "separation of concerns” paradigm to manage complexity, no longer seem to work. For example, fully verifying machine learning (ML) based perception processing subsystems in isolation might not be feasible. But when such perception processing is viewed in conjunction with the control components they feed into, full verification of the ML components might not be necessary to provide "global” correctness guarantees. Similarly, while high-level models of control components can be formally verified today, verifying software implementations synthesized from such models, especially when they are running on complex distributed architectures is more difficult. Building simulations and virtual prototypes are both difficult and expensive and still do not offer the kind of guarantees that are necessary. Along the same lines, while the real-time systems literature provides an array of formal methods for timing analysis and scheduling, few of them scale to the complexity found in real-life autonomous systems.
In summary, addressing the verification challenge for autonomous systems requires more holistic approaches that can "connect” different subsystems together for design, optimization, and verification. Such cross-layer verification requires design automation and suitable design tools. Do such tools and methods already exist? Which gaps in the availability of design automation tools are already recognized in the industry? What kinds of scientific challenges in developing such necessary tool support is recognized by the academic research community? And most fundamentally, does design automation indeed hold the key to achieving verifiable autonomy without overprovisioning resources?
The goal of this panel would be to discuss and debate on these questions. It will shed light on perspectives of the academic research community and also that of the industry. The outcome of this panel discussion would be an actionable research agenda for embedded systems and design automation researchers.
Event Type
Research Panel
TimeThursday, July 13th3:30pm - 5:30pm PDT
Location3014, 3rd Floor
Autonomous Systems