Considerations of Artificial Intelligence Safety Engineering for Unmanned Aircraft

Abstract

Unmanned aircraft systems promise to be useful for a multitude of applications such as cargo transport and disaster recovery. The research on increased autonomous decision-making capabilities is therefore rapidly growing and advancing. However, the safe use, certification, and airspace integration for unmanned aircraft in a broad fashion is still unclear. Standards for development and verification of manned aircraft are either only partially applicable or resulting safety and verification efforts are unrealistic in practice due to the higher level of autonomy required by unmanned aircraft. Machine learning techniques are hard to interpret for a human and their outcome is strongly dependent on the training data. This work presents the current certification practices in unmanned aviation in the context of autonomy and artificial intelligence. Specifically, the recently introduced categories of unmanned aircraft systems and the specific operation risk assessment are described, which provide means for flight permission not solely focusing on the aircraft but also incorporating the target operation. Exemplary, we show how the specific operation risk assessment might be used as an enabler for hard-to-certify techniques by taking the operation into account during system design.

Sharpening the Scythe of Technological Change: Socio-Technical Challenges of Autonomous and Adaptive Cyber-Physical Systems

“…This hypothesis is broken if AI is embedded in the software of the drone, for example by allowing a learning phase. Although specific drones do not require certification, the assessment of safety critical properties could be problematic by the introduction of AI and, in particular, of the machine learning algorithms, as discussed in [33].…”

“…However, the AI introduction and, in particular, the (on-line or off-line) learning phase overstretch the fundamental hypothesis on which the traditional certification process is based (i.e., the learning phases overstretch reproducibility of proofs that ensure the same system behavior under the same inputs). This situation is also problematic for the category of specific drones (which do not require certification [31,32]) to assess safety-related requirements [30] as discussed in [33]).…”

Thermal and Visual Tracking of Photovoltaic Plants for Autonomous UAV Inspection

“…As an aside, this is also confirmed by AI in aviation (particularly machine learning) still in its infancy. Even if AI is widespread in subdomains such as logistics and fuel consumption estimation [46,47], AI techniques in flight control are rarely found in real-world application scenarios.…”

Review on artificial intelligence techniques for improving representative air traffic management capability

“…Yu et al [53] integrated the underlying physics of aircraft dynamic systems into machine learning models, to reduce training costs, and for accurate prediction of flight trajectories. Wang et al [54] introduced a method that mapped the raw sensory data of unmanned aerial vehicle (UAV) into control signals, which enabled the UAVs to autonomously generate suitable trajectories in virtual large-scale complex environments. Schirmer et al [55] introduced current certification practices in unmanned aviation, supported by autonomous systems and AI, and demonstrated that it is possible to use specific operation assessment as an enabler for hard-to-certify techniques.…”