Flight Simulator Training Device

• FSTDs are ground-based training systems that emulate aircraft dynamics with varying fidelity, from full-motion simulators to VR-based devices.
• They integrate advanced motion cueing techniques like 6‑DoF Stewart platforms to reproduce realistic flight dynamics, crucial for upset training.
• Modern FSTDs combine immersive visual, haptic, and secure software interfaces, balancing high training fidelity with human factors and regulatory needs.

Searching arXiv for recent FSTD-related papers to ground the article. A Flight Simulator Training Device (FSTD) is a ground-based device used for pilot training that replicates the aircraft environment and flight dynamics with a defined level of fidelity. In the materials considered here, the term spans full-motion Stewart-platform simulators for upset training, fixed-base and VR-based pilot trainers, mixed-reality display concepts, security architectures for classified simulator systems, workload-adaptive training environments, and adjacent research platforms for UAV and autonomy training. Across these contexts, an FSTD is characterized not only by its airframe and visual models, but also by the coupling of control interfaces, motion or vestibular cueing, scenario management, system security, and human-factors constraints (Zhao et al., 14 Mar 2025, Perez et al., 21 Jul 2025).

1. Definition, scope, and device classes

Under EASA, and similarly FAA/ICAO, the main categories are FFS – Full Flight Simulator, FTD – Flight Training Device, and FNPT / other training devices. An FFS is described as a high-fidelity replica with full cockpit, wide-angle visual system, and usually a 6-DOF motion platform; an FTD is a fixed-base simulator with full or partial cockpit; and FNPT and related devices are procedural trainers, generic cockpits, or lower-fidelity systems (Perez et al., 21 Jul 2025).

The surveyed literature also describes FSTDs functionally rather than only regulatorily. A Stewart-platform FSTD is presented as a 6‑DoF motion-base simulator intended to reproduce pilot motion sensations during demanding upset-prevention and recovery scenarios (Zhao et al., 14 Mar 2025). VR-based trainers are described as fixed-base or part-task devices with immersive visuals and cockpit hardware, even when not explicitly mapped to EASA or FAA qualification levels (Weelden et al., 2023). Mixed-reality systems are treated as a new visual/display technology for FSTD-class devices rather than as a new category of FSTD, with the cockpit, controls, flight model, and motion base remaining subject to existing certification specifications (Perez et al., 21 Jul 2025).

The term also extends into research-oriented simulators that are not certified FSTDs in the regulatory sense but architecturally parallel them. The FPV UAV platform built around AirSim/Unreal Engine and Pixhawk is described as a non-certified, high-fidelity FPV UAV training and research simulator, and the Birdly-based embodied drone system is treated as a motion-enabled UAV training environment with full-body control and multimodal feedback (Xiao et al., 2024, Cherpillod et al., 2017). This suggests that, in research usage, “FSTD” often denotes a broader design space of ground-based flight training and evaluation devices, provided that the device reproduces aircraft behavior, control interaction, and operational constraints with some defined degree of realism.

2. Motion-base FSTDs and six-degree-of-freedom cueing

A central hardware configuration in modern motion FSTDs is the 6‑DoF Stewart platform. The Stewart mechanism is adopted because of high rigidity, high maneuverability, high strength-to-weight ratio, and the ability to generate six degrees of freedom motion comprising 3 translations and 3 rotations (Zhao et al., 14 Mar 2025). In the cited work, a representative FSTD motion envelope is defined by translational excursions $x,y:\pm 1.7\ \mathrm{m}$, $z:2.2\sim 3.8\ \mathrm{m}$; translational velocities $x,y:\pm 1.5\ \mathrm{m/s}$, $z:\pm 1\ \mathrm{m/s}$; translational accelerations $x,y:\pm 10\ \mathrm{m/s^2}$, $z:\pm 7\ \mathrm{m/s^2}$; rotations roll and pitch $\pm 25^\circ$, yaw $\pm 30^\circ$; angular velocities $\pm 30^\circ/s$; angular accelerations $\pm 200^\circ/s^2$; and leg lengths $z:2.2\sim 3.8\ \mathrm{m}$0 (Zhao et al., 14 Mar 2025).

Such motion systems are especially relevant to Upset Prevention and Recovery Training (UPRT), including turbulence, wing stall, and high-rate attitude changes, where rapid, large changes in specific force and angular velocity must be reproduced within a limited linear workspace (Zhao et al., 14 Mar 2025). The associated control problem is therefore not merely one of mechanical actuation; it is one of allocating limited workspace to the most perceptually salient cues while remaining inside hard actuator and geometry bounds.

The literature contrasts classical and model-based cueing approaches. The Classical Washout Filter (CWF) uses high-pass translational and rotational channels for short-term cues and low-pass plus tilt-coordination channels to emulate sustained accelerations via platform tilt. It is described as effective for “typical” maneuvers but limited in UPRT because it adapts only in the parameter domain, does not explicitly consider full platform kinematics or dynamic translational–rotational coupling, and does not directly embed a vestibular system model (Zhao et al., 14 Mar 2025). These limitations become acute when large accelerations and angular velocities change quickly and with large magnitude.

To address this, the cited work formulates motion cueing as a constrained optimal control problem over the actual Stewart platform and human perception models. Its integrated prediction model couples vestibular dynamics, Stewart-platform kinematics, and pilot-head coordinate transforms. The vestibular subsystem includes semicircular-canal dynamics for rotational perception,

$z:2.2\sim 3.8\ \mathrm{m}$1

otolith dynamics for translational perception,

$z:2.2\sim 3.8\ \mathrm{m}$2

and tilt coordination,

$z:2.2\sim 3.8\ \mathrm{m}$3

The platform kinematics relate pose $z:2.2\sim 3.8\ \mathrm{m}$4 to leg lengths $z:2.2\sim 3.8\ \mathrm{m}$5 through inverse kinematics and linearized leg-rate equations $z:2.2\sim 3.8\ \mathrm{m}$6 (Zhao et al., 14 Mar 2025). This is significant because the optimization is posed directly in terms of perceived cues rather than only raw accelerations.

The resulting motion cueing framework includes MPC with terminal constraints, MPC without terminal constraints, and a Switchable Model Predictive Control (S‑MPC) architecture. The S‑MPC system combines MPC with COTC, MPC without COTC, an Adaptive Weight Regulator (AWR), a Supervisory Controller (SC), and a Switch Mixer (SM) (Zhao et al., 14 Mar 2025). The supervisory controller monitors feasibility of the constrained MPC optimization and the magnitude of lateral and longitudinal tracking errors; it switches to the unconstrained MPC when the terminally constrained problem becomes infeasible near or outside the operating envelope, then returns to the high-accuracy mode when the system re-enters a small error neighborhood. Soft switching is implemented by time-varying blending to avoid jerks perceptible to the pilot.

The quantitative consequence is improved motion-cue fidelity in aggressive upset scenarios. In a horizontal stall case relevant to UPRT, the reported Average Absolute Scale (AAS) values are $z:2.2\sim 3.8\ \mathrm{m}$7 for S‑MPC, $z:2.2\sim 3.8\ \mathrm{m}$8 for MPC, and $z:2.2\sim 3.8\ \mathrm{m}$9 for CWF, with S‑MPC reported to improve performance by 42.34% over MPC and 65.30% over CWF with respect to the AAS index (Zhao et al., 14 Mar 2025). A plausible implication is that, for motion-based FSTDs intended for upset training, fidelity hinges on simultaneous modeling of vestibular perception, actuator constraints, and switching logic at the edge of the usable workspace.

3. Visual, mixed-reality, and haptic interfaces

Visual-system architecture is a major axis along which FSTDs vary. Traditional high-end helicopter FSTDs use collimated or dome projection systems with wide field of view and carefully calibrated geometry, often paired with six-axis motion bases (Perez et al., 21 Jul 2025). MR HMD-based FSTDs instead replace or augment the outside-world visual system with a video see-through or optical see-through head-mounted display while leaving cockpit, controls, flight model, and motion base under existing certification requirements (Perez et al., 21 Jul 2025).

The MR survey identifies specific performance thresholds anchored in EASA special conditions. Continuous cross-cockpit field of view is specified as $x,y:\pm 1.5\ \mathrm{m/s}$0 horizontally, $x,y:\pm 1.5\ \mathrm{m/s}$1 up, and $x,y:\pm 1.5\ \mathrm{m/s}$2 down, while total motion-to-photon latency is required to satisfy

$x,y:\pm 1.5\ \mathrm{m/s}$3

for FTD Level 3, with frame rate

$x,y:\pm 1.5\ \mathrm{m/s}$4

The hardware-side mitigation strategy prioritizes high-resolution, low-latency, low-jitter HMDs, accurate IPD calibration, and integrated eye-tracking with gaze-driven autofocus to reduce vergence–accommodation conflict (Perez et al., 21 Jul 2025). At the training-protocol level, the same survey recommends 20–30 minute MR blocks with breaks, posture guidance, habituation schedules, airflow, and standardized questionnaires such as SSQ, VRSQ, and NASA‑TLX (Perez et al., 21 Jul 2025).

VR-based FSTDs can also serve as research instruments for pilot-state estimation. In a VR-based PC‑7 simulator with cockpit mock-up, stick and pedals with control loading, and a Varjo Aero HMD, novice military pilots performed a speed-change maneuver from 180 KIAS to 110 KIAS while maintaining altitude, heading, and coordinated flight (Weelden et al., 2023). The study derived Pilot Inceptor Workload (PIW) from control-stick data. For each axis, Duty Cycle was defined as

$x,y:\pm 1.5\ \mathrm{m/s}$5

with $x,y:\pm 1.5\ \mathrm{m/s}$6 indicating movement above threshold, and Aggressiveness as

$x,y:\pm 1.5\ \mathrm{m/s}$7

A one-dimensional workload index was then defined as

$x,y:\pm 1.5\ \mathrm{m/s}$8

Longitudinal PIW was predictive of workload in a pitch-dominated speed-change task, with logistic-regression coefficient $x,y:\pm 1.5\ \mathrm{m/s}$9, 95% CI $z:\pm 1\ \mathrm{m/s}$0, $z:\pm 1\ \mathrm{m/s}$1, whereas lateral PIW was not predictive (Weelden et al., 2023). This establishes the FSTD as both a training environment and a measurement apparatus.

Another interface trajectory is the replacement or virtualization of cockpit controls. The mid-air haptics concept paper proposes a VR-based cockpit with no physical panels, using ultrasound phased arrays such as Ultrahaptics STRATOS Inspire. The cited hardware characteristics include 256 Murata MA40S4S transducers at 40 kHz, peak power consumption of about 80 W, haptic projection distance up to 70 cm, an interaction cone of about $z:\pm 1\ \mathrm{m/s}$2, spatial resolution of about 4 mm, and focal-point update rates up to 40 kHz (Girdler et al., 2020). Amplitude modulation and spatio-temporal modulation are used to produce “click,” “dial,” “line,” and “presence” sensations corresponding to virtual pushbuttons, rotary knobs, levers, and throttle strips. The same paper explicitly frames current FFS economics as a problem: certified FTDs without motion cost hundreds of thousands of dollars, while certified full-flight simulators with full motion cost tens of millions of dollars, and each FFS is limited to one aircraft type (Girdler et al., 2020). The proposed value of mid-air haptics is therefore software reconfigurability rather than force-fidelity equivalence.

At the low-cost end of the interface spectrum, a wearable haptic glove is used as a tangible controller for a Unity-based flight simulator. The glove uses a MEMS IMU for hand orientation, flex sensors for finger bending, and fingertip vibratory motors for event-triggered haptic feedback, with Kinect used for hand-position tracking (Foottit et al., 2016). The device does not reproduce aircraft control geometry or force loading, but it demonstrates that alternative embodied and vibrotactile interfaces can support intuitive control metaphors in flight-like virtual environments.

4. Adaptive, physiological, and human-factors-aware training

Recent FSTD research extends beyond aircraft-state fidelity into adaptation to the trainee’s internal state. A neuro-adaptive VR flight training system for the Pilatus PC‑7 integrates a cockpit mock-up with stick and pedals with control loading, a Varjo Aero HMD, headphones, and a wireless 32-channel EEG cap (Weelden et al., 9 Dec 2025). The EEG processing chain uses MNE-Python, a 3–35 Hz band-pass filter, a 50 Hz notch filter, automatic ICA, spectral power in theta, alpha, and beta bands, and the EEG Engagement Index

$z:\pm 1\ \mathrm{m/s}$3

A stacked classifier consisting of SVM, Random Forest, and Logistic Regression in the first layer and an SVM in the second layer was trained on prior VR-PC‑7 data to discriminate “Low” versus “High” workload (Weelden et al., 9 Dec 2025). Reported average offline performance was accuracy $z:\pm 1\ \mathrm{m/s}$4, F1 score $z:\pm 1\ \mathrm{m/s}$5, precision $z:\pm 1\ \mathrm{m/s}$6, and recall $z:\pm 1\ \mathrm{m/s}$7 (Weelden et al., 9 Dec 2025).

The adaptive logic changes simulator difficulty between 120-second trials. If EEG-estimated workload is “High,” the next trial’s difficulty is reduced by one level; if it is “Low,” difficulty is increased by one level (Weelden et al., 9 Dec 2025). Difficulty levels include clear weather, mist, heavy fog, attitude-indicator failure, and false horizon. In the reported experiment with 15 Royal Netherlands Air Force student pilots, there were no significant differences between adaptive and fixed-sequence conditions in NASA‑TLX, UES‑SF, SSQ, or an objective performance metric defined as

$z:\pm 1\ \mathrm{m/s}$8

However, in both conditions flight performance decreased as subjective workload increased, with Spearman $z:\pm 1\ \mathrm{m/s}$9 in the fixed condition and $x,y:\pm 10\ \mathrm{m/s^2}$0 in the adaptive condition (Weelden et al., 9 Dec 2025). Interview data indicated that, once briefed, 10 of 15 pilots preferred the adaptive condition. This suggests that neuro-adaptive FSTDs may alter perceived realism and engagement before they produce clear immediate performance gains.

Human-factors management is especially central in MR-based FSTDs. The MR survey organizes 18 drivers of cybersickness and related issues into internal, hardware, simulation-content, and ergonomics clusters. It highlights sensory conflict, latency and latency jitter, field of view, user susceptibility, exposure duration, visual complexity, vergence–accommodation conflict, IPD misalignment, HMD weight, head-movement frequency, posture, cognitive load, presence, and affective state (Perez et al., 21 Jul 2025). Several mitigation measures are classified as high priority and high feasibility in the surveyed literature: HMD choice, IPD calibration, preserving the real cockpit as a stable fixed reference, posture guidance, airflow, habituation, oculomotor training, and positive onboarding (Perez et al., 21 Jul 2025). By contrast, common VR sickness mitigations such as teleportation, reverse optical flow, and aggressive FOV restriction are described as poorly suited to aviation because they break realism and can conflict with EASA field-of-view requirements (Perez et al., 21 Jul 2025).

A common misconception is that increasing immersion alone improves training. The cited works instead indicate that immersion must be balanced against cybersickness, visual fatigue, ergonomic strain, and workload-management constraints. In this sense, an advanced FSTD is not simply a higher-resolution or wider-FOV device; it is a system whose display, control, adaptation, and session structure are jointly tuned to maintain pilot performance and training effectiveness.

5. Security, software infrastructure, and operational integration

An FSTD can also be treated as a sensitive information system rather than only as a flight-dynamics simulator. In a Department of Defense context, the Prometheus flight simulator program is described as exposing highly sensitive data, including tactics, mission profiles, training scenarios, and “top secret” materials, and therefore requiring a secure, auditable, DoD-compliant identity and access-control framework (Slaughter et al., 2011).

The proposed architecture places token and PKI authentication interfaces in a DMZ, while the Flight Simulator Application Server, back-office systems, directory services, and Oracle database remain on the internal network (Slaughter et al., 2011). Supported second factors include Public Key Infrastructure (PKI) with Common Access Cards (CAC) and token-based methods such as RSA SecurID, intended to meet NIST Level 4. The PKI process validates the certificate chain against a dedicated Certificate Authority and checks revocation status using OCSP and CRL. Conceptually, confidentiality and integrity are expressed using standard asymmetric cryptography and hash verification: $x,y:\pm 10\ \mathrm{m/s^2}$1

$x,y:\pm 10\ \mathrm{m/s^2}$2

Digital signatures provide non-repudiation: $x,y:\pm 10\ \mathrm{m/s^2}$3 The architecture also uses SSO, ACLs tied to user roles, weekly network audits, twice-daily scans, and incident-response procedures aligned with NIST SP 800‑61 (Slaughter et al., 2011).

This security framing broadens the meaning of FSTD infrastructure. In high-assurance environments, the simulator is embedded in a larger ecosystem of remote access, scheduling, training records, administrative controls, and auditability. A plausible implication is that FSTD fidelity cannot be reduced to cockpit and motion realism alone in classified or distributed training environments; confidentiality, integrity, availability, and traceability become part of the device’s operational specification.

Software modularity and interoperability also recur in adjacent simulation research. The FPV UAV trainer uses a Python/Kivy GUI, OpenCV, Pymavlink, sockets, and a Pixhawk-based HIL loop, while the stochastic autonomous fixed-wing simulator separates Earth model, aircraft model, flight physics, sensors, navigation, guidance, and control in a modular C++ implementation (Xiao et al., 2024, Gallo, 2023). The latter is organized around actual, sensed, observed, visual, and reference trajectories, with error metrics such as Flight Technical Error, Navigation System Error, and Total System Error. This suggests that FSTD software architecture is increasingly converging on explicit separation between physics truth, sensor emulation, estimation, control, and training-side scenario logic.

6. Extensions beyond conventional pilot trainers

Several cited systems sit outside classical certified FSTD categories yet illuminate where the concept is expanding. The open-source FPV UAV platform combines a physical F450 quadcopter, dual-view FPV, a Pixhawk flight controller, and an AirSim/Unreal Engine digital twin with hardware-in-the-loop. The digital twin reproduces a real test facility and supports tasks such as takeoff-hover-land, A-to-B flight, obstacle avoidance, and figure‑8 maneuvers (Xiao et al., 2024). Flight states, joystick signals, and images are timestamped and synchronized, while GPS trajectories are transformed from LLA to ECEF and then ENU for analysis: $x,y:\pm 10\ \mathrm{m/s^2}$4 In Task 1, for example, the physical system reports hover distance $x,y:\pm 10\ \mathrm{m/s^2}$5, height deviation $x,y:\pm 10\ \mathrm{m/s^2}$6, and time $x,y:\pm 10\ \mathrm{m/s^2}$7, while the digital twin reports $x,y:\pm 10\ \mathrm{m/s^2}$8, $x,y:\pm 10\ \mathrm{m/s^2}$9, and $z:\pm 7\ \mathrm{m/s^2}$0, respectively (Xiao et al., 2024). The platform is explicitly described as non-certified but functionally FSTD-like.

The Birdly system pushes this further toward embodied interaction. A commercial VR flight simulator is coupled to a real outdoor quadcopter so that full-body hand gestures drive fixed-wing-like flight through a mapping from desired pitch and roll to quadcopter velocity commands. The control transformation includes

$z:\pm 7\ \mathrm{m/s^2}$1

$z:\pm 7\ \mathrm{m/s^2}$2

$z:\pm 7\ \mathrm{m/s^2}$3

$z:\pm 7\ \mathrm{m/s^2}$4

$z:\pm 7\ \mathrm{m/s^2}$5

In a four-condition simulation study, Birdly + attitude mapping achieved average evaluation performance of $z:\pm 7\ \mathrm{m/s^2}$6, exceeding Birdly + angular velocity ($z:\pm 7\ \mathrm{m/s^2}$7), RC + angular velocity ($z:\pm 7\ \mathrm{m/s^2}$8), and RC + attitude ($z:\pm 7\ \mathrm{m/s^2}$9) (Cherpillod et al., 2017). This demonstrates that flight training devices can be organized around body-centric interaction and multimodal presence rather than around cockpit replication.

Autonomy-oriented simulators likewise broaden the scope of what an FSTD can train. A photorealistic quadrotor simulator couples Gaussian Splatting radiance fields to PyBullet dynamics and trains visual navigation policies using Liquid neural networks (Quach et al., 2024). The recurrent component is based on the Closed-form Continuous-time formulation

$\pm 25^\circ$0

with closed-form hidden-state update

$\pm 25^\circ$1

The policy outputs body-frame velocities and yaw rate,

$\pm 25^\circ$2

and is trained by imitation on structured single-target trajectories. In GS-to-real indoor transfer, the Liquid model reports total success rate $\pm 25^\circ$3, while in GS-to-real outdoor transfer it reports $\pm 25^\circ$4, whereas the LSTM baseline fails in the outdoor setting (Quach et al., 2024). Although not a pilot FSTD in the regulatory sense, it functions as a simulator-based training device for autonomous flight behaviors. A plausible implication is that the concept of “training device” is extending from human skill acquisition to the pretraining and validation of onboard autonomy.

7. Limitations, controversies, and future directions

Across the cited literature, several limitations recur. First, many results are simulation-based or prototype-based rather than obtained in certified operational devices. The Stewart-platform motion-cueing results are evaluated in simulation under UPRT-relevant scenarios (Zhao et al., 14 Mar 2025). The neuro-adaptive VR trainer reports no significant gains in workload or objective performance relative to a fixed sequence, despite pilot preference for the adaptive condition (Weelden et al., 9 Dec 2025). The workload-from-stick-input study involves only six novice military pilots and reports Tjur’s $\pm 25^\circ$5, which it explicitly characterizes as weak but non-zero (Weelden et al., 2023). The MR survey also notes that most evidence still comes from VR studies, with limited MR-specific data on long-term adaptation and transfer (Perez et al., 21 Jul 2025).

Second, fidelity is multidimensional and sometimes conflicting. Mid-air haptics promises software reconfigurability and reduced hardware cost, but current systems are vibrotactile rather than force-reflecting and therefore do not satisfy higher-level certification assumptions that controls, switches, and knobs physically replicate the aircraft (Girdler et al., 2020). MR HMDs can reduce infrastructure cost and increase immersion, but they introduce cybersickness, visual fatigue, and ergonomic strain that do not arise in the same way with collimated dome systems (Perez et al., 21 Jul 2025). Embodied and glove-based controllers may be intuitive and engaging, but they do not reproduce certified cockpit geometry, force-feel, or procedural workflows (Cherpillod et al., 2017, Foottit et al., 2016).

Third, model dependence remains a structural issue. The Stewart-platform controller depends on accurate vestibular and kinematic models (Zhao et al., 14 Mar 2025). The autonomy-transfer simulator depends on the fidelity of the GS scene, the alignment between physics and rendering, and the robustness of the network architecture (Quach et al., 2024). The open-source AGISim inertial-sensor simulator depends on JSBSim flight dynamics, lever-arm transforms, and configurable IMU error models,

$\pm 25^\circ$6

$\pm 25^\circ$7

and explicitly notes the absence of gimbal dynamics and more advanced stochastic processes (Kazemi et al., 2024). This suggests that future FSTDs will increasingly be judged not only by cockpit appearance or motion-base specifications, but by the validity of their integrated state-estimation, human-perception, and sensor models.

A final controversy concerns standardization. The materials describe clear EASA thresholds for MR HMD use and well-established FFS/FTD categories (Perez et al., 21 Jul 2025), but several emerging device types—mid-air-haptic cockpits, embodied UAV trainers, and neuro-adaptive VR systems—do not fit neatly into existing qualification classes (Girdler et al., 2020, Cherpillod et al., 2017, Weelden et al., 9 Dec 2025). This suggests that future FSTD standards may need to distinguish more explicitly between airframe/vehicle fidelity, cueing fidelity, interface fidelity, human-factors robustness, and adaptive or autonomy-related functions.

Taken together, the cited work portrays the FSTD as a converging technical system: a device in which motion cueing, cockpit and control embodiment, visual and mixed-reality presentation, physiological adaptation, software modularity, and information security are all treated as integral to training fidelity. The historical image of the FSTD as simply a cockpit on a motion base remains central, but the recent literature indicates a broader and more computationally integrated conception of what a training device is and what, exactly, it is meant to reproduce (Zhao et al., 14 Mar 2025, Perez et al., 21 Jul 2025).