qronOS Overview

• Publications
• People
• Theses

Related Projects

• RTSC
• KESO
• I4Copter

qronOS

Quality-aware Co-Design of Responsive Real-Time Control Systems

Selected Publications

Overview

When driving a car, a human pays full attention when required, but efficiently reduces the mental workload in less critical situations. On the other hand, current control strategies for autonomous cars always pay "full attention" and require precise timing. How can we make real-time control systems efficient and safe at the same time, just like a human can do it intuitively?

Timing plays a particularly important role in the solution for this problem: As control applications are particularly sensitive to timing variations, the Quality of Control (QoC) is degraded by varying execution conditions of the underlying real-time system. This is particularly relevant for modern, adaptive real-time systems, which introduce varying timing conditions for the sake of efficiency increases.

We aim to solve this design conflict by explicitly considering the relation between timing (Quality of Service, QoS) and the resulting Quality of Control both at design- and run-time. To do so, the QRONOS project is an interdisciplinary cooperation between the Chair of Automatic Control and the Chair of Distributed Systems and Operating Systems, ensuring that both real-time and control engineering aspects are correctly considered in the creation of a common framework.

Efficient Real-Time Control with Safety Guarantees (RTAS-WiP 2019)

At RTAS-WiP 2019 we presented the key concept of the QRONOS project: To achieve efficient average-case use of the real-time computing system, but at the same time provide hard worst-case safety guarantees, we propose a hybrid approach consisting of two parts: Efficient normal operation and a predictable safety fallback.

Normally, an optimistic operation mode is active in which the execution requirements (QoS) of the control application are reduced as far as the current situation (disturbance and physical state) permits. This mode is optimized for average-case efficiency and allows for high flexibility in the real-time system, including techniques such as mixed-criticality scheduling which provide higher efficiency at the cost of reduced temporal guarantees.

Because verification of the optimistic mode is impractically complex and significantly restricts the flexibility in real-time system design, a safety net is employed instead: A safety mechanism detects that an unsafe situation is imminent and avoids the situation by switching to a predictable safety mode. This safety mode provides verified worst-case behavior of the control system by providing predictable timing within small bounds.

• full paper will be available in the RTAS 2019 BP proceedings. (preprint)

Leveraging Machine Learning Techniques for Quality-Aware Real-Time Scheduling

The above mentioned optimistic execution mode involves the redistribution of resources between real-time tasks. One way to achieve this goal is to omit unnecessary control task invocations. That is task executions that would benefit the quality-of-control beyond a defined target value. In other words, control tasks should be omitted or deferred by the scheduling, when the current situation and QoC levels permit.

To still guarantee that the desired QoC level is not violated by a scheduling decision in the near future, we propose an anticipatory scheduling approach that takes past iterations as well as the current QoC state.

The solution to this problem is algorithmically complex and computationally expensive. Therefore, in QRONOS we leverage machine learning techniques to implement decision making at runtime as efficiently as possible. Based on a Linear Impulsive System (LIS) model that provides the ability to simulate the temporal evolution of the QoC in a noisy environment, we propose a machine-learning framework for decreasing the utilization associated with an existing control algorithm. Since the state follows non-linear trajectories, machine learning is employed to segment the state-space by means of small Multilayer Perceptrons (MLPs). The Supervised Learning Policy (SLP) policy shows excellent classification performance at a fraction of the cost that would be necessary for a normal model calculation. The video shows that our approach is able to maintain a certain QoC level while ensuring that future control steps will not violate the QoC aim.

Worst-Case QoC-Analysis (ARCH 2019)

At ARCH 2019 we presented the worst-case verification of control systems with uncertain timing as a benchmark for the verification of hybrid dynamical systems. The experimental results indicate that the problem is challenging and can currently only be solved for some cases, but is challenging in general.

• full paper will be available in the ARCH 2019 proceedings. (preprint)
• slides
• software and data (GPLv3): citable archive with source and output / newest sourcecode on GitHub

Stochastic QoC-Model (HSCC 2018)

At HSCC 2018 we presented an approach towards efficient computation of the Quality of Control in a time-varying stochastic sense (see the full paper and the slides).

The software, which passed the official repeatability evaluation, can be found on GitHub and is GPL3-licensed. The original repatability evaluation archive is also available for download. Re-use and collaboration is appreciated.

Full Publications List

People Involved in qronOS

Dr.-Ing. Peter Ulbrich

Dr.-Ing. Andreas Michalka

Maximilian Gaukler, M.Sc.

Dipl.-Ing. Tobias Klaus

Theses