Human as an Actuator Dynamic Model Identification

TL;DR

The paper develops a time-domain framework to identify a parametric human pilot model θ that, when combined with vehicle dynamics, yields an actuator-like representation of pilot behavior. Data from multiple simulator trials are fused through a constrained optimization that enforces dynamic consistency via a differentiation operator, assuming Gaussian measurement noise. A single θ is estimated across experiments by aggregating residuals with per-run dynamic constraints and stability/actuator-bounds, demonstrated on a quadrotor pitch-control task. The quadcopter experiment shows the estimated pilot model reproduces observed vehicle responses across different target distances, illustrating robustness to human variation and beneficial regularization from the vehicle model. The approach is extensible to multi-input, multi-output and nonlinear scenarios, broadening applicability to diverse piloted-vehicle operations.

Abstract

This paper presents a method for estimating parameters that form a general model for human pilot response for specific tasks. The human model is essential for the dynamic analysis of piloted vehicles. Data are generated on a simulator with multiple trials being incorporated to find the single model that best describes the data. The model is found entirely in the time domain by constructing a constrained optimization problem. This optimization problem implicitly represents the state of the underlying system, making it robust to natural variation in human responses. It is demonstrated by estimating the human response model for a position control task with a quadcopter drone.

Figures (5)

• Figure 1: An illustration of a height-velocity diagram used by helicopter pilots. This shows the unsafe flight regimes as shaded areas. The dynamics of the human pilot must be appended to the system to accurately calculate these diagrams.
• Figure 2: An example where the pilot model was estimated using the pilot commands directly. The resulting model applied to the system is shown in red. We can see in the figure that this fails to converge to the goal state. A pilot model fitted using the vehicle dynamics as a regularizer is shown in yellow and is observed converging to the goal position.
• Figure 3: The graphical depiction of the pilot model and how it forms a joint system with the vehicle.
• Figure 4: A screen capture of the simulation environment used for the example in Section \ref{['results']}. A marker representing the target position is plotted on screen and the pilot, starting from rest, maneuvers the vehicle (in the foreground) to stop directly on top of the target.
• Figure 5: The data that was recorded when the pilot was controlling the quadcopter and shown in blue. The green lines show goal positions. The estimated model's response to the same scenarios are shown in red for a 10m goal and shown in yellow for the 20m goal. The same single model was used for both the yellow and green goal states. This shows the quality of the model fit is robust to the natural variation from human controlled vehicles.