Periodic Reporting for period 1 - DARWIN (Digital Assistants for Reducing Workload & Increasing collaboratioN)
Reporting period: 2023-06-01 to 2024-12-31
A steady increase in conventional traffic which needs to be handled in the airspace safely and efficiently, increased demand for new pilots connected to a risk of pilot shortage occurring in the future, growing environmental concerns, and the emergence of new entrants into the airspace are the fundamental challenges that must be successfully addressed.
The advent of Artificial Intelligence (AI) offers potential to address these challenges, provided it can be implemented safely and in harmony with human operators.
The DARWIN (Digital Assistants for Reducing Workload and Increasing collaboratioN) project aims to support flights or parts of the flights operated by a single pilot, with the enabling technology validated on a commuter aircraft (CS-23) and subsequently used in the future as a steppingstone to large aircraft (CS-25) market.
The support to the single pilot shall be achieved by developing a Level 2 AI system (as defined by EASA) for Human-AI collaboration assisting the pilot in selected situations. This main objective is further split into three technical objectives:
• Develop & Validate Trustworthy Machine Reasoning AI Platform
• Develop & Validate Pilot State and Task Load Monitor
• Develop & Validate Human-AI Teaming technology
The trustworthy and certifiable AI platform provides explainable and reliable decision support and decision-making capabilities. The Pilot State Monitor can detect pilot drowsiness, sleep, and obvious incapacitation and pass this information to the systems capable of selecting a proper reaction. The Task Load Monitor determines the current and expected task load for pilots along the flight path, considering changes during the flight over time as well as foreseeable changes in the environment surrounding the flight path. Human-AI Teaming develops and validates innovative means of how to allocate tasks between the human pilot and a range of digital assistants in a way that keeps an acceptable level of the pilot’s workload, while keeping him or her in the loop and in charge of the aircraft.
DARWIN applies both logic- and knowledge-based (e.g. machine reasoning), as well as data-driven (machine-learning) methods and algorithms that are not yet directly applicable in the safety-critical cockpit environment. The project analyses regulatory gaps (missing regulation and guidance) and challenges (existing regulations and guidance) related to the applied methods and their software implementation. Wherever applicable, these activities are being aligned with ongoing EASA work in the AI field.
The advent of Artificial Intelligence (AI) offers potential to address these challenges, provided it can be implemented safely and in harmony with human operators.
The DARWIN (Digital Assistants for Reducing Workload and Increasing collaboratioN) project aims to support flights or parts of the flights operated by a single pilot, with the enabling technology validated on a commuter aircraft (CS-23) and subsequently used in the future as a steppingstone to large aircraft (CS-25) market.
The support to the single pilot shall be achieved by developing a Level 2 AI system (as defined by EASA) for Human-AI collaboration assisting the pilot in selected situations. This main objective is further split into three technical objectives:
• Develop & Validate Trustworthy Machine Reasoning AI Platform
• Develop & Validate Pilot State and Task Load Monitor
• Develop & Validate Human-AI Teaming technology
The trustworthy and certifiable AI platform provides explainable and reliable decision support and decision-making capabilities. The Pilot State Monitor can detect pilot drowsiness, sleep, and obvious incapacitation and pass this information to the systems capable of selecting a proper reaction. The Task Load Monitor determines the current and expected task load for pilots along the flight path, considering changes during the flight over time as well as foreseeable changes in the environment surrounding the flight path. Human-AI Teaming develops and validates innovative means of how to allocate tasks between the human pilot and a range of digital assistants in a way that keeps an acceptable level of the pilot’s workload, while keeping him or her in the loop and in charge of the aircraft.
DARWIN applies both logic- and knowledge-based (e.g. machine reasoning), as well as data-driven (machine-learning) methods and algorithms that are not yet directly applicable in the safety-critical cockpit environment. The project analyses regulatory gaps (missing regulation and guidance) and challenges (existing regulations and guidance) related to the applied methods and their software implementation. Wherever applicable, these activities are being aligned with ongoing EASA work in the AI field.
There are several highlights from the first half of the project. This involves developing the initial concept of operations for Human-AI Teaming, and analysing pilot tasks for upcoming single pilot operations. It also involves defining preliminary use cases to validate the project outcomes.
An interactive demonstrator was created to test the dynamic allocation and reallocation of tasks between human pilots and automation. It was validated with pilots and reached Technology Readiness Level 4 – “technology validated in lab”. First connection of the demonstrator to Pipistrel aircraft avionics was successfully tested in order to prepare for upcoming flight tests in 2025.
Main technical gaps have been identified on the Trustworthy Machine Reasoning Platform (TMRP) and successfully addressed. In addition, regulatory and standardization aspects of use of such AI system in aviation have been explored. TMRP has also been integrated into the interactive demonstrator and will be validated in upcoming validations.
Pilot State and Task Load Monitor has been split into two sub-objectives – Pilot State Monitor, and Task Load Monitor.
- Notable progress in the Pilot State Monitoring (PSM) includes the refinement of obvious incapacitation detection technology for laboratory and aircraft integration, with improvements in both accuracy, robustness, and computational efficiency. This monitor has been successfully validated with pilots and reached Technology Readiness Level 4.
- For the Task Load Monitor (TLM), the definition of task load for the project, based on the current state of scientific research, was finalised and made available to the project. Based on this, the TLM has been further developed and now communicates fully with the other integrated DARWIN systems. The task load can now be analysed and evaluated in four dimensions. Unexpected external events are taken into account. In addition, a Flight Management System (FMS) has been integrated to make the analyses even more precise. The overall system has been validated with airline pilots and has reached Technology Readiness Level 4.
An interactive demonstrator was created to test the dynamic allocation and reallocation of tasks between human pilots and automation. It was validated with pilots and reached Technology Readiness Level 4 – “technology validated in lab”. First connection of the demonstrator to Pipistrel aircraft avionics was successfully tested in order to prepare for upcoming flight tests in 2025.
Main technical gaps have been identified on the Trustworthy Machine Reasoning Platform (TMRP) and successfully addressed. In addition, regulatory and standardization aspects of use of such AI system in aviation have been explored. TMRP has also been integrated into the interactive demonstrator and will be validated in upcoming validations.
Pilot State and Task Load Monitor has been split into two sub-objectives – Pilot State Monitor, and Task Load Monitor.
- Notable progress in the Pilot State Monitoring (PSM) includes the refinement of obvious incapacitation detection technology for laboratory and aircraft integration, with improvements in both accuracy, robustness, and computational efficiency. This monitor has been successfully validated with pilots and reached Technology Readiness Level 4.
- For the Task Load Monitor (TLM), the definition of task load for the project, based on the current state of scientific research, was finalised and made available to the project. Based on this, the TLM has been further developed and now communicates fully with the other integrated DARWIN systems. The task load can now be analysed and evaluated in four dimensions. Unexpected external events are taken into account. In addition, a Flight Management System (FMS) has been integrated to make the analyses even more precise. The overall system has been validated with airline pilots and has reached Technology Readiness Level 4.
Since the project has deployed an agile approach, there are several interim results already available. An initial version of human-AI teaming concept of operations tailored for aerospace domain and aligned with EASA guidelines for trustworthy and reliable AI has been drafted, implemented through an interactive demonstrator, and validated with pilots. The results are now being integrated into a new version of the human-AI teaming concept. Upon completion, this work will lay a foundation and establish guidelines for effective interaction and collaboration between humans and AI.
The Trustworthy Machine Reasoning Platform (TMRP) has been further matured and successfully tested in several situations. The TMRP represents an important enabler for incorporating trustworthy and explainable AI to the cockpit. With this technology, the pilots will gain insights into the actions of the automation and the rationale behind them, thereby enhancing cooperation and trust with AI.
Additionally, a successful validation of obvious pilot incapacitation detection has been conducted. Once implemented in aircraft, this detector can further improve safety by e.g. facilitating responses to incapacitation incidents, especially during single pilot operations and extended minimum crew scenarios, where one pilot may be resting during the cruise phase of flight.
The Task Load Monitor (TLM) can already predict the task load along a planned route with a high degree of accuracy. To further improve the system, it is necessary to consider not only the new route, but also the impact of rescheduling on the overall system. Areas for improvement include understanding the dynamics of task load, in particular the effects of multitasking in high demand situations and the cognitive load of switching between unrelated tasks. Investigating these factors will improve task prediction and task allocation.
In addition, improving the temporal accuracy of task predictions is crucial. A more accurate determination of the temporal distribution of tasks in flight will enable better analysis of task distribution and management strategies. Further development of TLM requires research into the effects of multitasking, cross-domain task switching, and precise task timing. Addressing these challenges will improve task allocation and promote safer and more efficient flight operations.
The project is going to focus on the further maturation of these results, their integration into an aircraft, and validating them during a real flight.
The Trustworthy Machine Reasoning Platform (TMRP) has been further matured and successfully tested in several situations. The TMRP represents an important enabler for incorporating trustworthy and explainable AI to the cockpit. With this technology, the pilots will gain insights into the actions of the automation and the rationale behind them, thereby enhancing cooperation and trust with AI.
Additionally, a successful validation of obvious pilot incapacitation detection has been conducted. Once implemented in aircraft, this detector can further improve safety by e.g. facilitating responses to incapacitation incidents, especially during single pilot operations and extended minimum crew scenarios, where one pilot may be resting during the cruise phase of flight.
The Task Load Monitor (TLM) can already predict the task load along a planned route with a high degree of accuracy. To further improve the system, it is necessary to consider not only the new route, but also the impact of rescheduling on the overall system. Areas for improvement include understanding the dynamics of task load, in particular the effects of multitasking in high demand situations and the cognitive load of switching between unrelated tasks. Investigating these factors will improve task prediction and task allocation.
In addition, improving the temporal accuracy of task predictions is crucial. A more accurate determination of the temporal distribution of tasks in flight will enable better analysis of task distribution and management strategies. Further development of TLM requires research into the effects of multitasking, cross-domain task switching, and precise task timing. Addressing these challenges will improve task allocation and promote safer and more efficient flight operations.
The project is going to focus on the further maturation of these results, their integration into an aircraft, and validating them during a real flight.