Member of the Cybernetics Approach to Perception and Action research (CAPA) group and the Motion Perception and Simulation (MPSim) group.
Over the last 4 years I worked on my Phd Research project resulting in my thesis titled "Control-Theoretical Models of Feedforward in Manual Control". Below you find a summary of this research, identical to the summary text in the thesis. Where possible, links to articles and videos are included to make the summary better understandable for non-experts.
Being in control of a vehicle is part of everyday life for many people. Understanding how humans control a vehicle is especially important for the design of vehicles and their interfaces to the human controller. It allows engineers to design faster, safer, more comfortable, more energy efficient, more versatile, and thus better vehicles. Especially now, when automation enables us to support the human controller in every way imaginable, it is important to understand how the human controls and interacts with a vehicle. The human and the automation will dynamically share the control authority over the vehicle. Hence, the automation should (at least!) be designed around the human, but it would be much better if the automation behaves in a similar way to the control behavior of the human. If the automation behaves as a human controller, the human controller understands the
intentions of the automation better, which leads to a higher safety, increased comfort and ready acceptance.
The Human Controller (HC) is almost always in control of the vehicle to achieve a high-level goal. To achieve this high-level goal, the HC needs to perform a great number of smaller tasks in succession that are achieved by giving “control inputs” to the vehicle: moving the steering wheel, pressing the gas pedal, pulling the collective lever in a helicopter, turning a rotary knob, etc. To understand the relationship between the high-level goal and the low-level control inputs, it is helpful to distinguish between three types of behavior: skill-based, rule-based, and knowledge-based behavior. Knowledge-based behavior relates to complex decisions made by the human in order to achieve the high-level goal, such as those required to take the fastest route through a busy city during rush-hour. Rule-based behavior relates to simple actions performed in an “if-then-else” fashion, such as stopping for a traffic light if it lights red. Skill-based behavior relates to automatic sensori-motor patterns of behavior, such as steering left and right to stay within the lines of the road. While executing a ‘sensori-motor pattern’, the human continuously perceives certain signals from the environment through the senses, such as visually perceiving the distance to the side of the road, and acts by giving control inputs to the vehicle by moving the hands or feet. This thesis focuses on sensori-motor patterns executed during short, single maneuvers, such as a lane-change maneuver or a turn in a car; a sidestep, bob-up, or pedal turn maneuver in a helicopter; or a landing flare, take-off, or decrab maneuver in an aircraft.
In my PhD work, I studied the sensori-motor patterns of control behavior by means of “target-tracking and disturbance-rejection control tasks”. In such a task, the HC gives control inputs such that the vehicle tracks a particular reference path, the target, as accurately as possible. The vehicle (often called the system) is perturbed by disturbances and the HC is required to reject (attenuate) the resulting deviations of the system from the target. The HC can use closed-loop feedback, open-loop feedforward, or a combination of both.
In closed-loop feedback control, the HC senses and responds to the tracking error, i.e., the difference between the current output of the system and the desired output (the target). Every realistic control task involves disturbances, which can only be attenuated through feedback control, and thus it is likely that the HC uses feedback control. For good tracking performance, feedback control requires the HC to respond to the tracking error with a small time delay, but often the time delay is too large. Therefore, it is unlikely that the HC relies entirely on feedback control. In open-loop feedforward control, the commands given by the HC to the system are based on the target only; the HC does not compare the actual system output with the target. Feedforward control provides a much better tracking performance than feedback control, but it does require the HC to have extensive knowledge of the target and the system dynamics. The HC obtains knowledge of the target by visually perceiving it and by predicting the future course of the target. It is unlikely that the HC relies entirely on feedforward control, because a) the HC does not have perfect knowledge of the target and the system, and b) external disturbances are generally unknown and unpredictable. Thus, the HC likely uses a combination of feedforward and feedback.
The HC will use a pure feedback control strategy only if both the target and the disturbance are unpredictable and the HC can only perceive the tracking error from the display. Such tasks are extremely rare in the real world. Yet, almost all HC models describe the human as a pure feedback controller, but the important feedforward response received little attention. Therefore, the goal of this thesis is to obtain a fundamental understanding of feedforward in human manual control.
Based on the results of two initial studies, the following four objectives towards achieving the thesis goal were established.
The first objective was to develop a novel system identification procedure that allows for the objective identification of feedforward and feedback behavior in tracking tasks modeled after realistic control tasks. The two initial studies had shown that existing methods were unsuited for this purpose. The novel procedure successfully addressed the three central issues in system identification for manual control. First, the procedure does not require the user to make assumptions regarding the model structure and/or dynamics, which makes the results more objective than those obtained with previous methods. Second, the procedure explicitly prevents ‘false-positive’ feedforward identification: models that include a feedforward path in addition to a feedback path have more parameters and therefore more freedom to fit the data, resulting in a better fit even if a true feedforward response was not present. Hence, if the ‘best’ model is selected based on the quality of the fit alone, a ‘false-positive’ feedforward identification is possible. The procedure therefore imposes a penalty on model complexity, the weight of which is tuned based on Monte Carlo simulations. Third, the procedure is able to identify the correct HC dynamics from data containing high levels of human noise measured under closed-loop feedback conditions. The procedure was then successfully used to address the other three objectives of the thesis.
The second objective was to investigate how the HC adapts the feedforward dynamics to the system dynamics and the waveform shape of realistic target signals. First, it was found that the theoretically ideal feedforward dynamics are equal to the inverse of the system dynamics. For example, if the system dynamics are a single integrator, the ideal feedforward dynamics are a differentiator. From a number of human-in-the-loop tracking experiments, it was concluded that the HC utilizes feedforward dynamics that are indeed very similar to the inverse of the system dynamics. Deviations from the ideal dynamics are due to limitations in the perception, cognition, and action loop of the HC. These limitations can be modeled accurately by a gain, a time delay, and a low-pass filter. The HC was found to utilize a feedforward response with three different system dynamics (a single integrator, a second-order system, and a double integrator) and two target signal waveform shapes (consisting of either constant velocity ramp segments or constant acceleration parabola segments).
The third objective was to investigate how the subjective predictability of the target signal affects feedforward behavior. The central hypothesis of feedforward behavior states that the HC will develop a more optimal feedforward strategy easier if the target signal is more predictable. The predictability of a target signal is affected by many factors, here the predictability of a sum-of-sine target signal was investigated, by an objective system identification analysis, and subjects were asked to give a subjective rating of predictability. It was found that the feedforward gain was higher for signals rated more predictable, and that the feedforward time delay was close to zero for the most predictable signals, which suggests that subjects were indeed anticipating the future course of the target signal.
The fourth objective was to investigate how human feedforward interacts with other HC responses, primarily the feedback response on the system output in tasks that feature physical motion feedback. The HC can potentially use three control responses in a realistic control task in which physical motion feedback is present: a feedforward on the target, a feedback on the tracking error, and a feedback on the system output. It was expected that the best tracking performance is obtained if all three responses are used simultaneously. A theoretical analysis revealed that the feedforward dynamics should adapt to the presence of an output feedback response for the performance to be optimal. That is, the ideal feedforward path is not equal to the inverse system dynamics, but equal to the sum of the inverse system dynamics and the dynamics of the output feedback path. From a human-in-the-loop experiment it was concluded that subjects indeed utilized all three control strategies simultaneously, but that they respond with a significantly smaller gain to the system output if they are simultaneously tracking a predictable ramp target signal.
The following general conclusions were drawn from the research work:
The developed system identification procedure and the feedforward/feedback HC model are valuable tools for future research on feedforward control behavior. The novel system identification procedure enables the researcher to obtain an objective estimate of HC control dynamics in control tasks that were not studied before. The application of the procedure is not limited to the identification of feedforward, it can be used to identify many other types of human dynamics. The HC model enables the researcher to investigate how task performance depends on the feedforward model parameters through computer simulations, it helps in formulating hypotheses, allows for effective design of experiments, and enables the researcher to get a deeper understanding of control behavior adaptations through parameter estimation analyses. The predictability of the target signal is the main point that needs further research, after which multi-loop, multi-axes control tasks need to be addressed. Eventually, research will have to move away from tracking tasks and investigate manual control behavior in tasks with fewer constraints and thus more freedom to follow a self-chosen path.
My thesis demonstrated that feedforward is an essential part of human manual control behavior and should be accounted for in many human-machine applications. The state-of-the-art in manual control was advanced considerably; a fundamental understanding of feedforward in human manual control was obtained.
Frank Michiel Drop was born on June 20th, 1986, in Amsterdam, the Netherlands. From 1998 to 2004 he attended Oostvaarders College in Almere, where he obtained his VWO diploma.
In 2004 he enrolled as a student at the Faculty of Aerospace Engineering at Delft University of Technology (Delft, the Netherlands). In academic year 2007/08 he was full time Technical Manager at Formula Student (FS) Team Delft, managing a diverse group of sixty undergraduate students in designing, building, and racing a single seat formula racecar. The team competed in the 2008 FS UK and FS Germany engineering competitions against university teams from all over the world, resulting in a 2nd and 1st place overall, respectively. He learned that all vehicles, especially racecars, perform best when the abilities and limitations of the human are considered during the design process.
As part of his M.Sc. studies at the Control and Simulation division of Aerospace Engineering, he performed an internship at the Max Planck Institute (MPI) for Biological Cybernetics in Tübingen, Germany, on the modeling of pilot control behavior during helicopter roll-lateral side-steps. This inspired him to investigate feedforward control behavior in realistic manual control tasks for his M.Sc. thesis work. He obtained the M.Sc. degree (cum laude) in May 2011.
After working at DESDEMONA as a simulation engineer, Frank started his Ph.D. project at the Max Planck Institute for Biological Cybernetics in collaboration with the Control and Simulation division at TU Delft in September 2011. He investigated feedforward control behavior and system identification methods for manual control research, which resulted in his PhD thesis. Since June 2015, he has been working on collaborative projects between MPI and BMW AG, Munich, Germany, involving the conceptual design of a novel motion simulator to be used by BMW for the reproduction of dynamic car-driving maneuvers. He is currently working as a research scientist at MPI on simulator Motion Cueing Algorithms involving Model Predictive Control techniques, and continues to work on human control models and identification methods.
, , , and (September-2016) Objective Model Selection for Identifying the Human Feedforward Response in Manual Control
IEEE Transactions on Cybernetics Epub ahead.
, , , and (December-2013) Identification of the Feedforward Component of Manual Control in Tasks with Predictable Target Signals
IEEE Transactions on Cybernetics 43(6) 1936-1949.
Conference papers (6):
, , and (August-2016) Constraints in Identification of Multi-Loop Feedforward Human Control Models, 13th IFAC/IFIP/IFORS/IEA Symposium on Analysis, Design, and Evaluation of Human-Machine Systems (HMS 2016), Elsevier, Frankfurt a.M., Germany, IFAC-PapersOnLine, 49(19), 7-12.
, , and (August-2016) The Predictability of a Target Signal Affects Manual Feedforward Control, 13th IFAC/IFIP/IFORS/IEA Symposium on Analysis, Design, and Evaluation of Human-Machine Systems (HMS 2016), Elsevier, Frankfurt a.M., Germany, IFAC-PapersOnLine, 49(19), 177–182.
, , and (June-16-2016) Objective ARX Model Order Selection for Multi-Channel Human Operator Identification, AIAA Modeling and Simulation Technologies Conference: Held at the AIAA Aviation Forum 2016, Curran, Red Hook, NY, USA, 787-803.
, , , , and (May-2014) Subjective and Objective Metrics for the Evaluation of Motion Cueing Fidelity for a Roll-Lateral Reposition Maneuver, 70th American Helicopter Society International Annual Forum (AHS 2014), Curran, Red Hook, NY, USA, 1706-1720.
, , , and (May-2013) Feedforward and Feedback Control Behavior in Helicopter Pilots during a Lateral Reposition Task, 69th American Helicopter Society International Annual Forum (AHS 2013), Curran, Red Hook, NY, USA, 1797-1811.
, , , , and (October-2012) Identification of the Transition from Compensatory to Feedforward Behavior in Manual Control, IEEE International Conference on Systems, Man, and Cybernetics (SMC 2012), IEEE, Piscataway, NJ, USA, 2008-2013.