Which Network Delays Make People Sick? Measuring and Scheduling Around Motion-Weighted Staleness in Networked VR Simulators

Virtual laboratory simulators, such as VR driving rigs and surgical rehearsal environments, are moving from single wired workstations to setups where physics and rendering run on an edge server and stream to a headset. Any network then has to decide which updates to deliver first and which to drop. Networking researchers score those decisions with Age of Information, which measures how old the freshest delivered update is at the receiver. Cybersickness researchers know that the body does not care about age in the abstract. It reacts to mismatches between what the eyes see and what the inner ear feels, and it reacts much more to rotational mismatch than to linear mismatch. We propose to measure, with human participants, which features of a network's age process actually cause sickness. We expect it is not the mean age, and probably not the peak age alone, but how much rotational motion the user missed while the display was stale. We then build a scheduler that spends bandwidth to keep exactly that quantity small.

The central bet is simple. Two networks with the same average staleness can differ sharply in how much sickness they cause. The difference depends on whether their stale intervals coincide with fast turns or with straight, steady driving. If that holds, the right currency for sickness-aware transport is a motion-weighted error, not time. The claim is new because the only work linking age metrics to motion sickness, On the Ruin of Age of Information in Augmented Reality over Wireless Terahertz (THz) Networks, treats sickness as an assumed threshold on peak age. It never puts a person in the loop.

Core Claim

In a networked simulator where the headset already corrects its own head rotation locally, the sickness-relevant staleness is in simulated ego-motion: the virtual vehicle turning, braking, or the camera following an instrument. The headset cannot correct that motion because it does not know it. We claim that per-minute sickness ratings and involuntary head sway track a vestibularly weighted display error much better than mean age, peak age, or ruin probability. Let $\theta(t)$ and $x(t)$ be the true simulated camera orientation and position, and $U(t)$ the generation time of the freshest delivered state. The metric is

$$M = \frac{1}{T}\int_0^T \Big(\alpha\,\|\theta(t)-\theta(U(t))\| + \beta\,\|x(t)-x(U(t))\|\Big)\,dt,$$

where $T$ is the session length and $\alpha,\beta$ are weights fitted from human data. We predict $\alpha \gg \beta$. That prediction comes from Effects of Linear Visual-Vestibular Conflict on Presence, Perceived Scene Stability and Cybersickness in the Oculus Go and Oculus Quest, which found that even inverted linear head-tracking gain left half of 14 participants with zero sickness. The authors attribute this to the otoliths, the inner-ear organs that sense linear acceleration, which tolerate mismatch better than the semicircular canals that sense rotation. $M$ is a version of the "Age of Incorrect Information" idea that Guest Editorial Signal Processing for XR Communications and Systems highlights for XR, which asks how wrong the displayed state is rather than how old it is. Our version differs in three ways: it splits rotation from translation, its weights are fitted against human sickness, and it is tested as a predictor.

Research Program

The platform is DReyeVR, an open-source CARLA-based VR driving simulator that logs 6-degree-of-freedom head pose and 120 Hz gaze. Its authors report simulator sickness as an unmitigated limitation. We split it so that vehicle physics and scene state run on a server and reach the headset through a controllable queue with network emulation. The headset still renders locally and corrects head rotation locally, so any injected staleness affects only the simulated ego-motion. A second task serves as a control. It is an anatomy-inspection scene modeled on IMHOTEP, which Pfeiffer et al. ran locally with 70 of 77 users rating it comfortable. It has almost no ego-motion, so the theory predicts that network staleness should barely matter there.

The work has three parts. First, a controlled perception study injects age processes with matched mean age but different structure. The second part fits and compares candidate metrics. The third builds and tests a state-dependent scheduler. The scheduler decides whether to send or preempt an update based on the predicted growth of $M$, which it estimates from the server's knowledge of upcoming vehicle angular velocity. This addresses a limitation stated in Information Freshness for Timely Detection of Status Changes: age-optimal sampling there is state-independent, even though entropy-based "generalized age" tracks detection delay under state changes where plain age does not.

The work also involves a methodological choice. The target signal is a fast online measure, not only the post-session questionnaire. We record the Fast Motion Sickness scale once a minute and head dispersion (involuntary postural sway from the headset's motion sensor). A novel method for VR sickness reduction based on dynamic field of view processing reports that head dispersion correlates at R = 0.82 with balance-platform stability. We also run the kinematic sickness predictor from A Deep Cybersickness Predictor through Kinematic Data with Encoded Physiological Representation, pretrained on the Cybersickness Virtual Navigation Dataset. This predictor supplies a within-session estimate of sickness level, so that the tests are not limited to pre/post differences in the Simulator Sickness Questionnaire.

Baselines

The do-nothing reference is the local, unnetworked simulator with zero added age. It sets the floor sickness level that network effects add to. The naive networked reference is FCFS (first-come-first-served) queueing with zero-wait sampling, meaning a new update is generated the moment the previous one is sent. This is what an unmodified streaming setup does. Joint Status Sampling and Updating for Minimizing Age of Information in the Internet of Things shows zero-wait is often suboptimal for age, so beating it is necessary but not impressive.

The strongest existing queueing competitors are preemptive last-come-first-served (LCFS) and non-preemptive LCFS. In preemptive LCFS, a fresh update displaces the one being transmitted. On the Ruin of Age of Information in Augmented Reality over Wireless Terahertz (THz) Networks reports a 12% increase in the probability of keeping peak age below ruin thresholds under high user density with preemptive LCFS. It also reports 45% lower average peak age for FCFS as bandwidth grows. Non-preemptive LCFS is included because Age of Information: The Gamma Awakening shows it beats preemption under near-deterministic service times at high arrival rates, which matches a fixed-size state packet. The strongest learned competitor is an age-minimizing scheduler in the style of DPDS from Age-Based Scheduling for Mobile Edge Computing: A Deep Reinforcement Learning Approach. That paper reports average age about 4x lower than local-only and about 8x lower than pure offloading across 15 devices. We retrain it on our traffic with the same number of environment interactions as our scheduler and give both the same hyperparameter search budget.

All policies share the same bandwidth cap, packet sizes, emulated channel traces, and driving routes. Route order is counterbalanced within subjects. None of the cited papers reports a human sickness outcome for any queueing policy, so no reported result was found on Lacuna for human sickness under FCFS, LCFS, or DPDS scheduling. The comparison numbers will be ours.

Beating the baselines means lower per-minute sickness scores and head dispersion than the best of FCFS, both LCFS variants, and DPDS at equal bandwidth, with a pre-registered minimum effect size. The scheduler should achieve this even while its mean age is equal to or worse than DPDS's. That combination is what would show that age is the wrong target.

The key ablation sets $\alpha=\beta$ in the scheduler's objective. This removes the rotation-versus-translation weighting while keeping state dependence. A second ablation keeps the weights but schedules on a fixed timetable, which isolates the value of using upcoming motion.

For the metric-comparison study, the prediction baselines are mean age, mean peak age, ruin probability computed as in the ruin-theory paper, and unweighted display error. The relevant published number for the kinematic predictor is 97.8% four-class accuracy on the navigation dataset from A Deep Cybersickness Predictor through Kinematic Data with Encoded Physiological Representation. We will report how far that accuracy drops when the predictor runs on our networked driving data. We do not expect it to transfer intact.

Why not just…?

Why not just rely on the headset's local reprojection, which already hides network lag? This is the most serious objection. Reprojection repairs head rotation using the headset's own sensor, and our design deliberately leaves that in place. What reprojection cannot fix is simulated ego-motion, because the headset does not know the car is turning until the server says so. The IMHOTEP-style control task tests this objection directly. If network staleness causes as little sickness in driving as in static inspection, reprojection already solves the problem and the project stops at the measurement study. That would still be a useful negative result.

Why not just minimize mean latency or pick the best standard queue discipline? This would suffice if sickness tracked mean age. Study 1 tests that by holding mean age fixed while moving stale intervals onto or off high-rotation segments. The literature already suggests the answer is not simple. The ruin-theory paper shows FCFS and preemptive LCFS trade off average age against tail events, and the edge-scheduling paper shows latency-first scheduling lets older state age in the queue. If a standard discipline matches our scheduler on sickness at equal bandwidth, we will report that the weighting adds nothing to scheduling, even if it helps prediction.

Why not just detect sickness and narrow the field of view when needed? Dynamic field-of-view restriction reduced head dispersion by 37.05% in the Lim et al. study, with a floor at 60° below which discomfort rose. That makes it a strong, network-agnostic competitor. It is a remedy applied after the mismatch, and it costs immersion. Our scheduler instead prevents the mismatch at no visual cost. We will run field-of-view restriction alone, the scheduler alone, and both together. If field-of-view restriction alone matches the scheduler, the transport-level contribution is weak. If the two add up, they are complementary.

Why not just buy more bandwidth? The ruin-theory paper reports that wider terahertz bandwidth lowered average age but not the severity of rare ruin events. We also sweep the bandwidth cap. If the scheduler's advantage disappears above a modest cap, we will state that cap as the regime boundary.

Proposed Experiments

Study 1 is within-subject with about 40 participants. That is larger than Kim et al.'s 14, and the navigation dataset's 53-person cohort gives a feasibility anchor. Participants drive a fixed 12-minute urban route under five age conditions on separate days. The conditions are zero added age, constant moderate age, the same mean with rare heavy-tail spikes placed on straight segments, the same spikes placed on turns, and staleness applied only to rotation or only to translation. We deliberately recruit beyond the young-adult skew of the existing dataset, whose mean age is 26.3. The decisive check is whether spikes on turns produce more sickness than identical spikes on straights, and whether rotation-only staleness exceeds translation-only staleness. If neither difference appears, the core claim fails.

Study 2 fits $\alpha$ and $\beta$ on half the participants. It then compares, on held-out participants, how well each candidate metric predicts per-minute sickness scores and head dispersion, using explained variance and rank correlation. The metric must beat mean age, peak age, and ruin probability on held-out people to count.

Study 3 compares the scheduler against all baselines and ablations at three bandwidth caps and two channel-trace families, with a fresh participant group. The primary outcome is Simulator Sickness Questionnaire change plus the area under the per-minute sickness curve. Secondary outcomes are head dispersion, driving lane-keeping error as a measure of training fidelity, and early dropout. A 45-minute extended session in the strongest condition probes the prolonged-immersion degradation reported in Quantifying the effects of working in VR for one week. The proposal succeeds if the weighted scheduler lowers sickness relative to the best standard policy at equal bandwidth while not lowering mean age. That result would show that the right thing to keep fresh is the rotational motion the user is feeling, not the time since the last update.