HDSim: Multifaceted High-Fidelity Simulators

• HDSim is a term for diverse simulation systems that bridge idealized models with real operational feedback in aerospace, high-harmonic generation, and autonomous driving.
• Each framework employs domain-specific fidelity techniques—such as hardware-in-the-loop feedback, coupled nonlinear field propagation, and LLM-driven perception modulation—to enhance simulation realism.
• These high-fidelity simulators enable controllable, analyzable, and actionable experiments by integrating realistic dynamics with simplified computational models despite inherent limitations.

Searching arXiv for “HDSim” to confirm whether the term is used by multiple distinct systems and to ground the entry in current papers. HDSim is a name used for multiple simulation systems in distinct research domains rather than a single universally standardized software package. In the arXiv literature, the term denotes at least three unrelated frameworks: a hybrid docking simulator for hardware-in-the-loop emulation of spacecraft docking at the DLR European Proximity Operation Simulator (EPOS) (Zebenaya et al., 2014), a monolithic, open-source simulation program for gaseous high-harmonic generation (HHG) that couples microscopic and macroscopic physics in a self-consistent calculation (Schröder et al., 2 Sep 2025), and an LLM-assisted human-driver traffic simulation framework for autonomous-driving evaluation (Li et al., 23 Aug 2025). The commonality across these uses is methodological rather than disciplinary: each system is designed to bridge an abstract model and a realistic operational environment, so that simulator outputs are closer to the physical or behavioral phenomena being studied.

1. Terminological scope and cross-domain usage

The term HDSim is polysemous in current research usage. In aerospace hardware-in-the-loop research, it refers to a hybrid docking simulator that combines real robotic hardware, measured contact forces, and numerically simulated spacecraft motion (Zebenaya et al., 2014). In ultrafast and strong-field optics, it refers to a code for simulating high-harmonic generation in gases, with coupled propagation of the fundamental laser field and the generated harmonic field in cylindrical geometry (Schröder et al., 2 Sep 2025). In autonomous-driving research, it denotes a human-driver traffic simulation framework that combines cognitive theory with LLM assistance to generate stylized background traffic in platforms such as CARLA (Li et al., 23 Aug 2025).

This naming overlap is significant because the three systems embody different notions of realism. The docking version seeks realism through hardware force/torque feedback and delayed robot actuation (Zebenaya et al., 2014). The HHG version seeks realism by retaining both single-atom dipole physics and macroscopic propagation effects in one program (Schröder et al., 2 Sep 2025). The traffic version seeks realism through a hierarchical human driver style model and Perception-Mediated Behavior Influence (PMBI), in which style alters the driver’s perceived scene rather than directly prescribing low-level control (Li et al., 23 Aug 2025).

A plausible implication is that “HDSim” functions less as a field-specific acronym with a fixed meaning than as a reusable naming pattern for high-fidelity simulators that explicitly mediate between idealized models and operationally realistic conditions.

2. HDSim as a hybrid docking simulator

In (Zebenaya et al., 2014), HDSim is a hybrid docking simulator, defined as a hardware-in-the-loop simulator that includes a hardware element within a numerical simulation loop. It was developed at the DLR European Proximity Operation Simulator for verification and validation of docking phases in on-orbit servicing missions such as satellite capture, repair, refueling, and deorbiting. Its defining architectural feature is a closed loop on the measured docking force/torque: the real docking interface generates actual contact forces, the numerical model converts those forces into relative spacecraft motion, and industrial robots physically realize that motion.

The simulator combines three subsystems. A real-time computer simulates the free-floating dynamics of the chaser and target spacecraft. Two industrial robots execute the commanded relative motion in position control. A docking mock-up with a force/torque sensor measures the real contact load, which is then fed back into the numerical simulator (Zebenaya et al., 2014). The paper emphasizes that the robots are very stiff and have non-negligible controller delays; the reported average EPOS control-system delay is about 32 ms in the closed loop, composed of two approximately 16 ms delays. That delay is central because it can destabilize contact when impacts are fast and compliance is low.

The contact model assumes point contact, frictionless tangential behavior, normal-force dominance, and a linear spring-damper law,

$f = -k d - b \dot d .$

For the 3D case, the chaser rigid-body dynamics are written as

$m \ddot r_G = {}^G f,$

$J \dot \omega_{BG} = {}^B \tau .$

A reduced-order 3D model exploits the fact that only the local normal direction drives the contact force, reducing the order from 18 states to 12 states (Zebenaya et al., 2014). For analytical stability analysis, the paper further restricts the system to a 2D model with state variables

$\{y, v_y, z, v_z, \theta, \omega\},$

and includes explicit delay dependence in the penetration depth and torque expressions.

A central analytical result is that the penetration-depth dynamics can be isolated as a delayed second-order system,

$m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$

with effective mass

$m_a = \frac{m}{1+\frac{m(a\cos\alpha)^2}{J_x}} .$

The characteristic polynomial of the linearized delayed system factors into a translation mode and a penetration/contact mode,

$$\chi_h(s)= \left[ m s^2 \right] \left[ m_a s^2+ e^{-sh} ( b s + k ) \right] .$$

The stability analysis shows that higher stiffness reduces delay margin, higher damping increases delay margin, more mass generally helps stability, and the stability boundary depends on the robot delay $h$ (Zebenaya et al., 2014).

The paper supplements pole analysis with a passivity observer based on measured input/output signals. The observed energy $\Delta E$ is computed from measured and commanded force/torque and velocity signals; $\Delta E<0$ denotes passive behavior, $m \ddot r_G = {}^G f,$0 lossless behavior, and $m \ddot r_G = {}^G f,$1 active behavior (Zebenaya et al., 2014). This is particularly useful in 3D, where closed-form predictive analysis is more difficult.

A major engineering addition is a passive compliance device attached to the chaser side. Its effective stiffness in the normal direction is

$m \ddot r_G = {}^G f,$2

and the simulator combines real compliance with virtual software stiffness and damping via

$m \ddot r_G = {}^G f,$3

This hybrid contact model enables tuning of impact behavior without redesigning the hardware interface (Zebenaya et al., 2014).

Experimental results support the analysis. In 1D tests with a chaser approach speed of 20 mm/s, mass about 63 kg, and delay 16 ms, increasing damping moved the system from unstable to stable: $m \ddot r_G = {}^G f,$4 gave $m \ddot r_G = {}^G f,$5, $m \ddot r_G = {}^G f,$6 gave $m \ddot r_G = {}^G f,$7, $m \ddot r_G = {}^G f,$8 gave $m \ddot r_G = {}^G f,$9, and $J \dot \omega_{BG} = {}^B \tau .$0 gave $J \dot \omega_{BG} = {}^B \tau .$1, where

$J \dot \omega_{BG} = {}^B \tau .$2

is the coefficient of restitution (Zebenaya et al., 2014). In 3D tests with spacecraft masses of 3000 kg each, principal inertias of 500 kg·m$J \dot \omega_{BG} = {}^B \tau .$3, compliance stiffness 4000 N/m, and optional virtual damping 40 Ns/m along the chaser body $J \dot \omega_{BG} = {}^B \tau .$4-axis, the damped case exhibited less rebound and more passive behavior than the undamped case (Zebenaya et al., 2014).

3. HDSim as a high-harmonic generation simulator

In (Schröder et al., 2 Sep 2025), HDSim is a monolithic, open-source simulation program for modeling gaseous HHG while keeping the microscopic and macroscopic physics together in one self-consistent calculation. The motivation is that HHG source performance depends on the coupled effects of laser propagation, gas dispersion, ionization, phase matching, absorption, and atomic response, especially under realistic conditions such as high pressure, tight focusing, long wavelengths, few-cycle pulses, and ionization-dominated regimes.

At its core, HDSim solves coupled propagation equations for the fundamental and harmonic fields in cylindrical geometry. The propagation equation is written as

$J \dot \omega_{BG} = {}^B \tau .$5

where $J \dot \omega_{BG} = {}^B \tau .$6 denotes the fundamental and $J \dot \omega_{BG} = {}^B \tau .$7 the harmonic field (Schröder et al., 2 Sep 2025). The fields are propagated separately but in lockstep, and are coupled through nonlinear source terms computed from the evolving laser pulse.

The microscopic harmonic emission is computed with Lewenstein’s strong-field model,

$J \dot \omega_{BG} = {}^B \tau .$8

with stationary momentum $J \dot \omega_{BG} = {}^B \tau .$9, action $\{y, v_y, z, v_z, \theta, \omega\},$0, and a hydrogenic dipole matrix element $\{y, v_y, z, v_z, \theta, \omega\},$1 as given in the paper (Schröder et al., 2 Sep 2025). The macroscopic harmonic polarization is then

$\{y, v_y, z, v_z, \theta, \omega\},$2

embedding the single-atom response in the evolving neutral fraction and gas density.

On the driver-propagation side, HDSim includes Kerr self-focusing and self-phase modulation through

$\{y, v_y, z, v_z, \theta, \omega\},$3

and plasma effects through a current density with free-electron and ionization-loss terms,

$\{y, v_y, z, v_z, \theta, \omega\},$4

Ionization rates are taken from the ADK tunneling model, with optional nonadiabatic corrections, and the ionic populations obey coupled rate equations (Schröder et al., 2 Sep 2025). This architecture is explicitly designed to capture strong propagation reshaping in high-intensity, long-wavelength regimes relevant to soft-x-ray generation.

The numerical implementation uses a Crank–Nicholson integrator for the propagation equation and a stepsize-adaptive explicit fourth-order Runge–Kutta method for the ionization equations. The software is implemented in C++, using Armadillo for array and linear-algebra operations and FFTW3 for fast Fourier transforms. It is not currently parallelized and is therefore intended for workstation-scale Linux machines rather than distributed runs (Schröder et al., 2 Sep 2025).

Configuration is text-based and designed for automation. The fundamental can be specified as a tabulated spectrum with spectral phase or by a Gaussian model with central frequency, bandwidth, amplitude, and polynomial spectral phase coefficients. Multiple Gaussian spectral components can be supplied simultaneously, enabling multicolor driving fields. The gas density profile $\{y, v_y, z, v_z, \theta, \omega\},$5 is defined by linearly interpolated position–pressure ratios $\{y, v_y, z, v_z, \theta, \omega\},$6, and dispersion is handled by pressure-scaled refractive indices: Sellmeier-type formulae for the fundamental and interpolated tabulated x-ray optical constants for the harmonics (Schröder et al., 2 Sep 2025).

The program supports several practical approximations. The Lewenstein integral may be restricted to shorter time windows, such as one optical cycle, for significant speed-up. If the fundamental varies slowly with $\{y, v_y, z, v_z, \theta, \omega\},$7, the atomic dipole may be computed only every $\{y, v_y, z, v_z, \theta, \omega\},$8-th plane and interpolated between planes (Schröder et al., 2 Sep 2025). These are presented as acceleration knobs that preserve essential physics while reducing runtime.

Validation examples include phase-matching scans in a short helium gas cell moved through focus, isolated attosecond pulse synthesis via amplitude gating, two-color driving fields with even harmonics, soft-x-ray HHG in high-pressure helium with strong propagation reshaping, and coherence-length mapping for source design (Schröder et al., 2 Sep 2025). The paper uses coherence-length analysis based on

$\{y, v_y, z, v_z, \theta, \omega\},$9

$m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$0

and

$m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$1

which makes the code useful not only for full spectral prediction but also for phase-matching diagnostics and source optimization (Schröder et al., 2 Sep 2025).

The paper is explicit about limitations. HDSim assumes a Gaussian beam profile, neglects harmonic back-action on the fundamental, is not parallelized, and does not yet include Laguerre-Gaussian beams. It is aimed at gaseous HHG and is rooted in strong-field approximation physics rather than full ab initio TDSE propagation for every atom (Schröder et al., 2 Sep 2025).

4. HDSim as a human-driver traffic simulator

In (Li et al., 23 Aug 2025), HDSim is an LLM-based human-like traffic simulation framework for self-driving tests. Its purpose is to make simulated background traffic behave more like real human traffic, on the premise that realistic surrounding vehicles are necessary for exposing safety-critical failures in autonomous-driving systems. The paper argues that existing background-traffic solutions are often too rule-based, too homogeneous, or too tied to low-level controls to capture diverse human behaviors.

The framework combines cognitive theory with LLM assistance and has two principal components: a hierarchical human driver style model and Perception-Mediated Behavior Influence (PMBI) (Li et al., 23 Aug 2025). The style model separates baseline driving competence from style-related effects. It consists of an inner Driving Capability Layer (DCL) and three outer Style Influence Layers (SILs):

• DCL: baseline driving competence.
• L1 – Personality Influence Layer: enduring traits such as aggressiveness or cautiousness.
• L2 – Physiological Influence Layer: temporary states such as fatigue or intoxication.
• L3 – Attentional Influence Layer: transient attention-related effects such as distraction.

The conceptual claim is that style influences perception and perception shapes behavior. PMBI implements this by allowing an LLM to modify the BEV observation input to a background autonomous-driving model, rather than letting the LLM directly choose throttle, brake, or steering commands (Li et al., 23 Aug 2025). The paper describes the resulting discrepancy between objective and perceived scene as an “illusion gap.” The underlying AD model remains the decision-maker, but it acts on style-biased perception.

The LLM is used in two stages: it first generates a style description from a style label or seed, and then translates that description into a policy set

$m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$2

where $m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$3, $m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$4, and $m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$5 are the personality, physiological, and attentional policies, respectively (Li et al., 23 Aug 2025). To turn those policies into executable transformations, the LLM is provided with 16 predefined perception modulation APIs, documentation for the APIs, and a retrieval-augmented set of 62 handcrafted scripts. These APIs can modify motion, spatial properties, temporal cues such as traffic lights, and structural properties such as lane alignment (Li et al., 23 Aug 2025).

The paper specifies distinct temporal update rules for the three layers. $m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$6 is translated once and kept fixed, $m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$7 is periodically updated, and $m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$8 is triggered randomly. Temporal consistency between two translations in the same policy layer is maintained following the Weber–Fechner Law and optic-flow sensitivity theory (Li et al., 23 Aug 2025).

The procedural PMBI loop includes obtaining the current simulator context, generating a style description $m_a \ddot d(t) + b \dot d(t-h) + k d(t-h) = 0,$9, translating policies, identifying objects in the BEV input, mapping the policies to API scripts, adjusting the BEV image, and feeding the modified input to the background AD model $m_a = \frac{m}{1+\frac{m(a\cos\alpha)^2}{J_x}} .$0 to produce the next decision $m_a = \frac{m}{1+\frac{m(a\cos\alpha)^2}{J_x}} .$1 (Li et al., 23 Aug 2025). This makes the system non-intrusive and model-agnostic: the paper states that existing simulators can replace rule-based or non-responsive background vehicles with HDSim-controlled ones without retraining the underlying AD model.

Experiments are performed in CARLA 0.9.10, Town05, with 30 concurrent SimAgents per run using the Roach expert policy (Li et al., 23 Aug 2025). Evaluated styles include aggressive and cautious behavior for L1, drunk and fatigued behavior for L2, and distracted behavior for L3. L2 is updated every 2000 simulation steps, and L3 is triggered stochastically by a Poisson process with arrival rate 0.064. Each experiment uses 10 routes, repeated three times (Li et al., 23 Aug 2025).

The system uses LLaMA 3.1 locally on an A800 GPU for low-latency validation and GPT-4o-mini for high-level reasoning and influence-script generation. Rendering is performed on one NVIDIA RTX 4090, while distributed simulation/runtime uses eight NVIDIA 2080 Ti GPUs (Li et al., 23 Aug 2025). Evaluation uses CARLA Leaderboard v1 metrics: Driving Score (DS) and Route Compliance (RC).

The paper’s central empirical claim is that stylized traffic reveals hidden weaknesses in autonomous-driving systems more effectively than conventional traffic. It reports that HDSim improves detection of safety-critical failures by up to 68% compared to standard testing, with one strong example being a 67.6% performance drop for AIM in aggressive traffic (Li et al., 23 Aug 2025). The affected AD models include InterFuser, TFPP, AIM, VAD, ST-P3, and LMDrive. The paper also associates simulated failures with real-world accident patterns via three case studies—aggressive cut-in, fatigued late reaction, and distracted missed-hazard collisions—aligned with NHTSA case IDs 2005009501684, 2005041508481, and 2007045403168 (Li et al., 23 Aug 2025).

For realism, HDSim is compared against CARLA’s parameter-based style module and ProSim. Using trajectories from the INTERACTION dataset annotated by five human drivers, the paper reports mean F1-scores across RF, SVM, and KNN classifiers of 97.8 for aggressive, 98.1 for cautious, 45.0 for distracted, and 72.7 for fatigued styles (Li et al., 23 Aug 2025). These exceed the baselines by an average of 23.3% over CARLA and 21.3% over ProSim, and the simulated distributions exhibit lower Wasserstein distances than ProSim for speed (38.16% lower), acceleration (27.42% lower), and heading (14.4% lower) (Li et al., 23 Aug 2025).

The paper also reports that the system scales approximately linearly with the number of agents, that local LLM memory usage is only 7.15 GB, that LLM reasoning is triggered only 3–6 times per route, and that overall runtime remains comparable to standard rule-based CARLA simulations (Li et al., 23 Aug 2025). This suggests that the framework is intended not merely as a proof of concept but as a practical multi-agent testing tool.

5. Comparative methodological structure

Despite their disciplinary separation, the three HDSim systems share a common structural pattern: each is designed to correct a mismatch between simplified simulation and operational reality.

In the docking simulator, the mismatch lies between rigid industrial robots and the compliant, delayed, contact-sensitive behavior of spacecraft docking. HDSim addresses this through a hybrid loop that combines real force sensing, virtual dynamics, passive compliance, and virtual damping (Zebenaya et al., 2014). In the HHG simulator, the mismatch lies between analytic treatments of isolated aspects of HHG and realistic source conditions in which microscopic dipole response and macroscopic propagation are inseparable. HDSim addresses this by solving coupled field-propagation equations while embedding Lewenstein single-atom emission in an evolving dispersive, ionizing medium (Schröder et al., 2 Sep 2025). In the traffic simulator, the mismatch lies between simplistic background traffic and the semantically rich, temporally varying behavior of human drivers. HDSim addresses this through layered style modeling and perception editing rather than direct low-level actuation (Li et al., 23 Aug 2025).

A plausible implication is that the name “HDSim” has converged on a family resemblance: it tends to denote simulators in which fidelity is achieved by explicitly modeling the interface between an idealized internal model and a realistic external process, whether that interface is contact, propagation, or perception.

6. Limitations, interpretive cautions, and research significance

Each HDSim variant is explicit about its limits. The hybrid docking simulator relies on simplifying assumptions such as stationary target, frictionless contact, point contact, and a linear spring-damper model, and its 3D analytical stability treatment was not fully developed in the paper (Zebenaya et al., 2014). The HHG simulator assumes a Gaussian beam profile, neglects harmonic back-action on the fundamental, is not parallelized, and is confined to gaseous HHG within strong-field approximation physics (Schröder et al., 2 Sep 2025). The traffic simulator depends on predefined perception APIs and scripts, uses periodic or stochastic style updates rather than a fully continuous cognitive model, is demonstrated in CARLA using BEV representations, and still incurs overhead relative to rule-based traffic (Li et al., 23 Aug 2025).

The significance of these systems therefore lies less in claiming exhaustive realism than in making realism controllable, analyzable, and experimentally actionable. The docking HDSim provides a framework for verification and validation of on-orbit servicing contact dynamics under delay and compliance constraints (Zebenaya et al., 2014). The HHG HDSim offers a reusable, open-source platform for HHG source design, phase-matching analysis, isolated attosecond pulse studies, and soft-x-ray optimization (Schröder et al., 2 Sep 2025). The traffic HDSim provides a plug-and-play mechanism for generating heterogeneous, interpretable human-like traffic that more effectively exposes safety-critical failures in self-driving systems (Li et al., 23 Aug 2025).

Taken together, these works indicate that HDSim is best understood not as a single simulator but as a recurrent label for high-fidelity simulation frameworks that couple abstract models to physically or behaviorally realistic feedback mechanisms.