The 5% Threshold: What Driving Simulators Can't (and Can) Tell Us About How Trucks Feel
The Question Every Truck Manufacturer Is Trying to Answer
Picture this: you're a test driver at a truck factory, and your job is to feel whether a new driveline feels right. Does the truck respond eagerly when you floor the accelerator? Does it surge forward smoothly, without hesitation or judder? Does the torque build feel satisfying, or does something feel off? You know good driveability when you feel it. But here's the problem—figuring out whether a new driveline has good driveability has always required building an actual truck first.
That's expensive. That's slow. And if you discover a problem late in development, fixing it can cost millions.
The alternative is a driving simulator. Instead of waiting for a physical prototype, engineers could test driveability with a virtual truck—a mathematical model running on a computer. The driver sits in a simulator cab, sees a virtual road on screens, and feels virtual accelerations through a moving platform. If this worked, manufacturers could catch driveability problems early, when they're still cheap to fix.
But there's a catch. Driving simulators can't perfectly recreate the motion of a real vehicle. The platform has limited travel. The cabin can't tilt infinitely. And most importantly, the question of whether a driver can actually perceive fine differences in acceleration through the simulator's motion cues has never been settled. If drivers can't feel subtle differences in the simulator, then the simulator can't be used to evaluate driveability.
That question—can professional drivers perceive small differences in longitudinal acceleration in a moving-base simulator?—is exactly what a team of researchers from the Swedish National Road and Transport Research Institute (VTI) and Linköping University set out to answer. Their results, published on arXiv, suggest not only that this is possible, but that the specific way the simulator maps vehicle motion to platform motion matters less than you might expect. And that's good news for the future of virtual prototyping in heavy vehicle development.
The Science: Building a Better Driving Simulator
The researchers worked at VTI's driving simulator, a specialized facility in Sweden. Unlike the hexapod-based motion platforms you might see in flight simulator training centers—those six-legged robots that can tilt in any direction—VTI's simulator uses a separate system for rendering translational acceleration. The platform sits on a long linear track, 7.5 meters of actual physical travel that can be used to reproduce the surge and sway of a moving vehicle.
This distinction matters. A hexapod simulator uses the same actuators for all motions—surge, sway, pitch, roll—which means there's competition for limited workspace. The VTI simulator can dedicate its linear rail specifically to horizontal acceleration, giving it a maximum travel of 7.5 meters, a maximum speed of 3.75 meters per second, and a maximum acceleration of 8.0 meters per second squared. The cabin can also tilt in pitch and roll to simulate sustained accelerations, and a vibration table reproduces vehicle dynamics.
The vehicle model itself was a six-degree-of-freedom multi-body dynamics model of a 6x4 tractor and semi-trailer. For this study, the researchers created a simplified electric driveline model that captured the essential behavior of a real truck in a full-throttle tip-in test—the kind of launch maneuver that excites the primary drivetrain dynamics including torque build-up, delay, jerk, and driveline oscillations.
The key variable in this experiment was something called the motion-cueing algorithm, or MCA. An MCA is the software that decides how to translate the infinite motion of a virtual vehicle into the finite motion of a real simulator platform. It has to make trade-offs: Should it render acceleration at full strength, even if that means running out of track space quickly? Should it tilt the cabin to simulate sustained acceleration, even though tilting can feel different from true acceleration? Should it scale down the acceleration to keep everything smooth, or prioritize authenticity over comfort?
Different applications call for different tuning. A racing simulator might prioritize high-g feel and responsiveness. A comfort study might prioritize smooth,缓缓 tilting. And for tip-in testing—a quick acceleration from standstill that lasts just a few seconds—yet another approach might be optimal.
The researchers examined three variants of the classical MCA, which works by splitting vehicle acceleration into low-frequency and high-frequency components. The low-frequency part is rendered by tilting the cabin (your inner ear can't distinguish a slow tilt from sustained acceleration), while the high-frequency part—sudden starts and stops—uses actual platform translation.
The three variants they tested were:
Unscaled Tip-In (UTI): Tuned specifically for short tip-in tests. The acceleration isn't scaled down at all. The cut-off frequency is set so the simulator uses the full 7.5-meter rail during a three-second full-throttle tip-in with 1.4 m/s² peak acceleration. When pre-positioned near the start of the track, it maximizes the motion experience.
Downscaled Tip-In (DTI): Also tuned for tip-in tests, but the acceleration is scaled down by a factor of 0.7. This allows longer tip-in maneuvers (up to five seconds) before the platform reaches the end of the rail. It trades some motion fidelity for more usable time.
Downscaled Generic (DG): The most general-purpose variant. It uses cabin tilt extensively to render sustained acceleration, which allows it to recreate a constant 1.4 m/s² acceleration indefinitely without exceeding workspace limits. The cut-off frequency is tuned differently—based on the steady-state response to a step input rather than the transient response of a tip-in.
Mean JND by Motion Cueing Variant
| Label | Value |
|---|---|
| UTI | 5.9 |
| DTI | 5.8 |
| DG | 6.4 |
The Experiment: Testing Perception Under Controlled Conditions
To measure how well drivers could perceive acceleration differences, the researchers used a classic psychophysical technique: the just-noticeable difference, or JND. The JND is the smallest change in a stimulus that a person can reliably detect—in this case, the smallest percentage change in longitudinal acceleration that a driver can perceive between two launches.
The method was a two-alternative forced choice (2AFC) task. Each trial presented drivers with two consecutive tip-in maneuvers. The first was always a standard stimulus at 1.4 m/s² peak acceleration—the reference. The second was a comparison stimulus, either slightly higher or slightly lower in acceleration. The driver had to say whether the comparison felt higher or lower than the standard. No guessing allowed; they had to pick one.
The researchers used a weighted staircase procedure to efficiently home in on each driver's threshold. Rather than testing many fixed difference levels, the procedure adjusted the difficulty after each response. Get it right, and the next trial gets slightly harder. Get it wrong, and it gets easier. The step sizes were tuned to target the threshold where drivers would answer correctly about 67% of the time—above pure chance (50%) but not so high as to require an impractical number of trials.
For practical reasons, only five absolute difference levels were used: 4%, 6%, 8%, 10%, and 12%. Each participant completed 20 trials per motion cueing variant, for a total of 60 trials per person.
To analyze the results, the researchers fitted psychometric functions to each participant's responses using a generalized linear model with a logistic link function—essentially, a mathematical curve that describes the probability of a "higher" response at each acceleration difference. The point where this curve crosses 50% probability is called the point of subjective equality (PSE)—the difference at which a driver perceives the two stimuli as equal. The JND was calculated as the average of the distance from PSE to the 75% point (upper JND) and from the 25% point to PSE (lower JND).
The participants were the secret weapon here. Ten professional driveability experts at a heavy truck manufacturer—all with prior experience of battery-electric heavy trucks similar to the modeled vehicle—served as test subjects. These weren't undergraduate psychology students paid $20 for course credit. These were the same people who evaluate driveability in real trucks. If they can perceive differences in the simulator, that's meaningful for real-world applications.
One participant was unable to complete the study due to simulator sickness, leaving nine for the final analysis.
In addition to the JND experiment, the researchers conducted a subjective comparison study. Each participant compared each pair of MCA variants by performing multiple tip-in maneuvers with one, then the other, and then rating whether the second was better or worse. This wasn't about detection thresholds—about whether the variants felt subjectively different in quality.
Point of Subjective Equality by Driver
| Label | Value |
|---|---|
| Driver 1 | -0.9 |
| Driver 2 | -3.3 |
| Driver 3 | -2.3 |
| Driver 4 | -0.6 |
| Driver 5 | -1.8 |
| Driver 6 | 1.4 |
| Driver 7 | -2.1 |
| Driver 8 | -3.5 |
What They Found: Consistency Across Algorithms
The headline result: there was no significant difference in JND between the three motion cueing variants. The mean JND across all participants and all MCA variants was 5.4%. In other words, drivers could reliably perceive an acceleration difference of about 5-6% in the simulator, regardless of which MCA was used.
To test this formally, the researchers ran paired t-tests comparing each variant. The mean difference in JND between DTI and UTI was -0.165 percentage points, with a p-value of 0.55—way above the 0.05 threshold for significance. The comparison between DG and UTI yielded similar results. The null hypothesis—that the mean JNDs were equal—could not be rejected.
This was somewhat surprising, but also reassuring. It suggests that for short tip-in maneuvers, the specific tuning of the motion cueing algorithm doesn't dramatically affect a driver's ability to perceive acceleration. The UTI variant, which uses the most translational motion and no scaling, produced similar JNDs to the DTI variant, which scales down acceleration by 30%. And both tip-in-tuned variants performed similarly to the more general DG approach, which relies more heavily on cabin tilt.
Individual variation was substantial. Looking at the per-participant JND estimates, some drivers could detect differences as small as 1.5% (Subject 3 with DTI) while others needed 14% (Subject 5 with UTI). The standard deviation across participants was 4.4% for UTI, 2.3% for DTI, and 7.2% for DG. DTI showed the tightest consistency across individuals; DG showed the widest spread.
Driver 1: JND Estimates Across Variants
| Label | Value |
|---|---|
| UTI | 4.1 |
| DTI | 6.9 |
| DG | 12 |
The point of subjective equality told an interesting story. Across all participants, the mean PSE was -1.9%, meaning drivers systematically perceived the second stimulus in each pair as higher in acceleration than the first, even when the two were identical. This is a known phenomenon in psychophysics called "sequential effect" or " expectation shift"—the brain recalibrates its reference based on the first stimulus. It's not a flaw in the experiment so much as a feature of human perception that researchers now need to account for.
In the subjective comparison, something else emerged: participants preferred the tip-in-tuned variants. When asked to compare DG to the other two variants, most participants rated the general-purpose algorithm as worse or much worse than UTI or DTI. This suggests that while detection ability (JND) was similar across variants, the subjective feel of the motion was not—and drivers noticed.
Motion sickness remained low throughout. None of the nine participants who completed the study reported symptoms greater than 7 on the Fast Motion Sickness scale, which ranges from 0 to 20. The researchers had taken precautions: pre-positioning the simulator near the start of the track to maximize workspace, limiting return deceleration to 0.2 m/s², and capping cabin tilt speed at 3 degrees per second (below most people's perception threshold). These choices paid off.
Why This Changes Things: The Path to Virtual Prototyping
For decades, the automotive and heavy vehicle industries have relied on physical prototypes for driveability assessment. You build it, you drive it, you feel it, you decide if it's good enough. This process works, but it's expensive and slow. A new driveline might take five years to develop; discovering at year four that the tip-in feel is unacceptable means either expensive redesigns or compromised quality.
The vision—evaluate driveability before a physical truck exists—is compelling. Engineers could test dozens of driveline configurations virtually, running simulations in days that would take months with physical prototypes. They could explore the trade-off space systematically, answering questions like: "Would a slightly stiffer engine mount improve tip-in feel, or would it introduce new problems?"
But for this vision to work, simulators must faithfully reproduce human perception. If drivers can't perceive the same differences in the simulator that they would in a real truck, the simulator is useless for driveability evaluation. Previous research had established that acceleration JNDs in real vehicles and simulators are roughly comparable—around 0.03 to 0.13 m/s² depending on the reference acceleration. But nobody had systematically tested how different motion cueing algorithms affect JND for tip-in maneuvers.
The VTI study provides the first evidence that, at least for short tip-in tests, the answer is: not much. The specific tuning of the MCA—whether you scale acceleration, how aggressively you use cabin tilt, where you set the cut-off frequency—doesn't significantly affect a driver's ability to detect acceleration differences. This is good news for simulator fidelity. It means manufacturers have flexibility in how they configure their motion cueing without worrying that one approach is blinding drivers to real driveability issues.
The 5.4% mean JND gives us a concrete threshold to work with. For a tip-in with a peak acceleration of 1.4 m/s², drivers can reliably detect changes of about 0.076 m/s². This aligns well with previous studies. Baumgartner et al. reported JNDs ranging from 0.03 m/s² at 0.346 m/s² reference to 0.059 m/s² at 1.384 m/s². Menig found JNDs of 0.071 m/s² at 1.448 m/s². The VTI result—0.076 m/s² at 1.4 m/s²—fits comfortably in this range.
The Weber fraction perspective makes this even clearer. Weber's law states that the ratio of JND to reference stimulus is approximately constant over a range of intensities. For accelerations above about 1.5 m/s², this ratio is roughly constant at around 5-10%. The VTI result of 5.4% is consistent with this. It means that if you're evaluating a driveline change that alters tip-in acceleration by less than 5%, most drivers probably won't notice in a real truck either. The simulator isn't more sensitive or less sensitive than reality—it's calibrated about the same way human perception is.
This has implications for how simulators can be used in development. If you know that drivers can detect 5% changes, you can design test protocols accordingly. You might test whether a new torque control algorithm changes tip-in feel by more or less than 5%. You might run comparative evaluations where one configuration is clearly better (10%+ difference) and trust that the ranking reflects real-world perception. Or you might be skeptical of any result showing only a 2-3% difference, knowing that such changes are below perceptual threshold.
The subjective findings add another layer. Even though drivers could detect acceleration differences equally well across all three MCA variants, they had preferences. The tip-in-tuned variants (UTI and DTI) were preferred over the general-purpose DG algorithm. This suggests that while detection sensitivity is similar, the qualitative feel of the motion matters for acceptance. A simulator used for driveability evaluation should probably be tuned for the specific maneuvers being tested, not for general-purpose use. The motion cueing doesn't affect what drivers can perceive, but it affects what they feel about the experience.
What This Opens Up: The Next Steps
The study is careful to acknowledge its limitations. Nine participants is not a huge sample, though for a specialized study with professional test drivers, it's reasonable. The study used only one type of maneuver—full-throttle tip-in from standstill. Different results might emerge for other driveability tests, such as gear shift quality, driveline shuffle during deceleration, or low-speed maneuvering.
The researchers didn't compare simulator results to real-vehicle results directly. They cite prior work showing that JNDs are similar between the two contexts, but a within-subjects comparison—same drivers evaluating both the simulator and a real truck—would strengthen the evidence. Such a study would be logistically challenging (you'd need access to a prototype and a test track) but would provide crucial validation.
The motion cueing variants were all based on the classical split-filter approach. More advanced algorithms exist—adaptive algorithms that adjust scaling based on predicted future motion, washout filters that prioritize different motion components, or optimization-based approaches that minimize perceptual error. Whether these newer methods yield different JNDs remains unknown.
One question the study raises but doesn't fully answer is why the tip-in-tuned variants were preferred even though detection sensitivity was equivalent. The researchers speculate that the general-purpose DG variant, which relies more heavily on cabin tilt, may introduce subtle false cues—moments where the driver feels something in the motion that doesn't match the visual scene. These false cues might not affect the ability to detect acceleration differences, but they could make the experience feel less natural or trustworthy. Understanding this trade-off between detection sensitivity and motion quality is an important direction for future research.
There's also the matter of learning and adaptation. All participants in this study were experienced test drivers with years of real-world driving experience. Would naive drivers, or drivers with less experience, show the same JNDs? Would the results change if drivers had extensive simulator experience, building mental models of how simulator motion relates to vehicle motion? These questions matter if simulators are to be used with broader populations, not just expert evaluators.
The practical implications, though, are clear. For manufacturers considering driving simulators for driveability assessment, the VTI study suggests that this is viable. The 5.4% JND gives a baseline for designing test protocols. The finding that MCA tuning doesn't dramatically affect detection suggests flexibility in implementation. And the preference for tip-in-tuned variants suggests that matching motion cueing to the specific application matters for user experience, even if it doesn't affect perceptual sensitivity.
The road from research finding to industry practice is long. Simulators are expensive—VTI's facility with its 7.5-meter track is not something every manufacturer can build. Integrating virtual driveability testing into development workflows requires changes to process, validation against existing methods, and cultural shifts in how decisions are made. But the physics and psychophysics are now clearer. A simulator doesn't have to perfectly recreate reality to be useful for driveability evaluation. It just has to stay within perceptual bounds—and those bounds are quantifiable.
The Bigger Picture: Accelerating the Future of Heavy Transport
Heavy trucks are undergoing a revolution. Electric drivetrains, autonomous systems, new materials, and sophisticated torque control algorithms are changing what's possible in vehicle dynamics. Electric motors can deliver torque instantly, precisely, and continuously—eliminating the lag and surge of traditional transmissions—but they also introduce new challenges in driveline damping, regenerative braking coordination, and one-pedal driving feel.
Evaluating these new drivetrains using traditional methods is becoming harder. Physical prototypes are more expensive. Development cycles are shorter. The number of configurations to evaluate is larger. If simulators can reliably capture driveability differences, they could become essential tools for navigating this complexity.
The VTI study is a small piece of a much larger puzzle, but it's a necessary piece. Before you can use a tool, you need to understand its limits. Understanding how human perception interacts with motion cueing algorithms is foundational knowledge. This study adds to that foundation, showing that for tip-in maneuvers, the specific approach to motion cueing matters less than you might expect—and that professional drivers can perceive clinically relevant differences in a well-designed simulator.
The 5.4% JND isn't just a number. It's a threshold between noticeable and unnoticeable, between a driveability problem that matters and one that doesn't. For engineers working on the next generation of heavy trucks, knowing that threshold—and knowing how to create it reliably in a simulator—could be the difference between catching problems early and living with them forever.
The future of heavy transport will be built virtually as well as physically. This study brings that future a little closer.