Let u∈U denote a candidate maneuver for the autonomous vehicle (AV) in a given driving scenario. This maneuver u may represent a trajectory segment (a sequence of (x,y,v,t) states) or a discrete control sequence (steering and acceleration inputs) over a finite horizon.

Let Ri(u) denote the predicted risk for road-user group i under maneuver u. Risk is defined here as a normalized harm proxy, typically computed as the integral of collision probability weighted by the expected severity of the collision (often parameterized by delta-V or kinetic energy transfer). We consider three distinct road-user groups:(1)i∈{Ego, Third, VRU}where:

Ego: Refers to the ego vehicle occupants.

The group-wise risk vector for a single scenario is thus expressed as:(2)R(u)=[REgo(u),RThird(u),RVRU(u)]⊤

This vector R(u) represents the distribution of group-wise risk induced by maneuver u and serves as the quantity optimized under a specified set of moral principles.

For reference, the “Ethical-Orig” baseline [1] combines Utilitarianism, Egalitarianism, Maximin, and Responsibility into a single scalar risk cost:(3)JRisk(u)=wBJB(u)+wEJE(u)+wMJM(u)−wRJR(u)here, wB,wE,wM,wR≥0 are scalar trade-off weights that determine the relative importance of each principle. 1.

Utilitarian/Bayesian Term (JB)

The utilitarian term JB(u) seeks to minimize the total aggregate risk. It is typically expressed as the average risk across all relevant actors:(4)JB(u)=1|SR|∑i∈SRRi(u)where SR denotes the set of relevant road users included in the risk evaluation for the scenario, and |SR| is its cardinality used for scale normalization. Minimizing JB aligns with the principle of efficiency—reducing the total expected harm regardless of who suffers it. 2.

Equality Term (JE)

The equality term JE(u) penalizes unequal risk allocation across groups, reflecting egalitarian values. A common instantiation uses pairwise absolute differences, e.g., (5)JE(u)∝∑i,j|Ri(u)−Rj(u)|

This term increases with the disparity among {Ri(u)}. While theoretically sound, unnormalized equality metrics can lead to “leveling down”, where the system prefers a scenario where everyone faces high risk (low variance) over one where some are safe and others are slightly at risk (high variance). 3.

Maximin (Worst-Off) Term (JM)

Based on Rawlsian principles of justice, specifically the Difference Principle, JM(u) focuses on the outcome of the least advantaged group. It implements a minimax strategy (minimizing the maximum risk):(6)JM(u)=maxiRi(u)

This term discourages maneuvers that impose extreme risk on any single group, serving as a robust safety floor. 4.

Responsibility Discount Term (JR)

The term JR(u) captures responsibility-sensitive risk discounting. In the baseline formulation Eq. (3), it enters as a subtractive term −wRJR(u). The intended effect is to reduce the protection weight for parties deemed responsible for the hazardous situation (e.g., a pedestrian jaywalking). If JR represents the “risk attributable to responsible parties,” subtracting it lowers the total cost, making the maneuver more attractive. 5.

Critique of the Baseline

In practice, the component terms {JB,JE,JM,JR} can exhibit incompatible scales. JB might be dominated by high-probability low-severity events, while JM is driven by rare high-severity spikes. This scale mismatch complicates weight tuning, often leading to solutions where one term dominates the others entirely. Moreover, the subtractive responsibility term introduces incentive distortions. By subtracting wRJR(u), the optimizer can effectively “earn points” by finding trajectories where high risk coincides with high responsibility. This reduces the penalty for harm rather than the harm itself, potentially encouraging aggressive behaviors toward non-compliant road users that conflict with the legal doctrine of “last clear chance.”

To address scale incompatibility and perverse incentives, the Ethical-Improved objective internalizes responsibility weighting, normalizes fairness terms, and introduces explicit VRU tail-risk regularization. The improved objective is defined as:(7)JRiskImp(u)=wB(1|SR|∑i∈SRδi(u)Ri(u))+wEJ^E(u)+wMJ^M(u)+λTmax{0,RVRU(u)−τ}2+λρmax{0,RVRU(u)−ρREgo(u)}2

The role and motivation of each component introduced in Eq. (7) are summarized below. 1.

Bounded Responsibility Discount Factor (δi)

Instead of subtracting responsibility as an external bonus, responsibility is implemented as a bounded discount factor δi(u) applied directly to the risk Ri(u) within the utilitarian sum:(8)δi(u)=max{ε,1−ri(u)},ε>0where ri(u)∈[0,1] denotes the responsibility score assigned to group i under maneuver u.

ri(u)=1 indicates full responsibility (e.g., a pedestrian jumping into traffic).

This formulation discounts the risk assigned to responsible parties but prevents the weight from ever reaching zero. The lower bound ε is crucial; it ensures that even for a fully responsible agent, the AV retains a non-zero incentive to avoid collision. This aligns with ethical frameworks that value human life intrinsically, regardless of fault, and legal standards that require minimizing harm even to negligent actors. By embedding responsibility inside the utilitarian aggregation, the subtractive “bonus” distortion is removed: cost reduction requires reducing weighted risk rather than exploiting responsibility assignments. 2.

Normalized Fairness Terms (J^E,J^M)

To ensure the trade-off weights wE and wM behave consistently across varying traffic densities and speeds, normalized fairness terms J^E(u) and J^M(u) are used, scaled to comparable magnitudes. A standard min–max normalization is:(9)J^(u)=J(u)−JminJmax−Jmin+ηwhere Jmin and Jmax are estimated from a representative set of training scenarios, and η>0 is a small constant for numerical stability (∼10−12). Alternatively, quantile normalization can be used to provide robustness against outliers. This normalization makes the weights wE and wM more interpretable as tuning parameters across heterogeneous scenarios. 3.

Explicit VRU Tail Regularization (λT)

This term addresses the “tyranny of the average” by explicitly penalizing extreme VRU risk via a squared hinge loss:(10)PenaltyT=λTmax{0,RVRU(u)−τ}2where the hinge operator is defined as h(x)=max{0,x}. The squared hinge form yields a quadratic penalty once the threshold is exceeded and provides a continuous gradient at the activation point, which is advantageous for gradient-based solvers.   •

Threshold τ: τ is a tolerance threshold, typically set to a high quantile of the VRU risk distribution (e.g., q∈[0.90,0.99]).

The final term imposes a relational constraint between VRU and ego risk:(11)Penaltyρ=λρmax{0,RVRU(u)−ρREgo(u)}2where ρ>0 is typically set near unity (e.g., ρ≈1).

This term discourages solutions in which VRU risk systematically exceeds ego risk, i.e., RVRU(u)>ρREgo(u). It structurally prevents the planner from achieving ego safety by disproportionately offloading tail risk onto VRUs, encoding a protective stance toward vulnerable parties.

Overall, Eq. (7) preserves the interpretability of the baseline multi-principle structure while explicitly targeting tail outcomes. Responsibility is incorporated as an internal weighting (avoiding subtractive incentives), fairness terms are scale-aligned via normalization, and the hinge- and ratio-based penalties suppress extreme VRU risk and persistent VRU-over-ego tail dominance. 5.

Parameter Interpretation and Calibration

The parameters in the Ethical-Improved objective are treated as policy-calibration parameters rather than universally optimal constants. The tail threshold τ defines the risk level above which VRU exposure is treated as an ethically salient tail event. For the planner-objective results reported in Section 4, τ is fixed at the empirical q0.95 VRU-risk quantile. The sensitivity analysis in Section 4.8 evaluates q_0.90, q_0.95, q_0.975, and q_0.99. The lower-bound parameter ε prevents responsibility weighting from eliminating the planner’s duty of care, while ρ controls the acceptable VRU/Ego tail-risk ratio. The numerical stabilizer η is used only to avoid division by zero in normalized fairness terms and does not encode an ethical preference.

In the simulator, the responsibility score ri(u) is assigned from scenario-level right-of-way ambiguity, occlusion exposure, and crossing-intent uncertainty. It is not used to absolve the planner from avoiding harm; instead, it adjusts the internal weighting of predicted risk through δi(u)=max{ε,1−ri(u)}. Thus, even road users with higher responsibility retain a nonzero safety weight.

To support IoV cooperative driving, the improved objective is extended from a purely trajectory-centric formulation to a communication-aware objective JRiskImp(u;zt), where zt records V2X availability, packet delivery ratio, age of information, RSU coverage, occlusion-aware perception confidence, and edge inference delay. The communication state does not replace local sensing; instead, it adjusts the normalized group-wise risk proxy and adds a penalty for stale, missing, or delayed cooperative perception. In occluded scenes, the VRU tail threshold and VRU/Ego ratio cap can be tightened so that tail penalties activate earlier when infrastructure support is unreliable.

The communication-aware objective is specified explicitly as Eq. (12), where the communication penalty is defined in Eq. (13). (12)JIoV(u,zt)=JImproved(u;R~(zt))+λcPcomm(zt)where (13)Pcomm(zt)=wp(1−PDR)+wamax(0,AoI−AoImax)+wr(1−RSUcov)+wq(1−qperc)+wdmax(0,dedge−dmax)here, a lower packet delivery ratio (PDR), a larger age of information (AoI), unavailable or partial RSU coverage (RSUcov), lower perception confidence qperc, and a larger edge inference delay dedge all increase Pcomm(zt). These communication degradations also inflate the communication-conditioned risk proxy R~i for occluded VRU states, so stale or missing cooperative perception makes the planner more conservative and activates the VRU tail-risk and VRU/Ego ratio penalties earlier, while local sensing remains available at every planning cycle.

We compare four distinct decision-making methods, differing only in the objective function minimized during trajectory planning: 1.

Standard: a safety-oriented baseline that minimizes the unregularized risk surrogate over the planning horizon. It should be interpreted as a planner without explicit fairness or tail regularization, not as an oracle for minimizing realized harm totals.

We used a paired-scenario evaluation methodology. All methods were evaluated on the same scenario set so that differences in outcomes can be attributed to decision logic and communication conditions rather than to stochastic variation in the environment.

Sample Size: We evaluated N = 2000 paired scenarios.

Scenario proportions: The 2000 scenarios were sampled as 25% unprotected left turns, 25% occluded crossings, 25% right-of-way ambiguity, and 25% heterogeneous crowds (500 paired cases per class).

Scenario Composition: The scenarios were curated to include diverse urban driving situations, including: ∘

Unprotected Left Turns: The AV must navigate a gap in oncoming traffic while watching for pedestrians on the crosswalk.

For each scenario n and each method, we logged the following telemetry:

Group-Wise Risk Vector: Rn=[REgo(n),RThird(n),RVRU(n)]⊤, representing the estimated risk levels for the ego vehicle, third-party road users, and vulnerable road users, respectively.

For the IoV extension, each paired scenario also logs end-to-end (E2E) latency, packet delivery ratio, message-drop sensitivity, age of information (AoI), communication overhead, edge delay, RSU coverage, and occlusion-specific safety metrics. A lightweight communication-aware simulator samples baseline-specific communication states, evaluates candidate maneuvers {brake, yield, balanced, proceed, evade}, and reports both ethical and IoV-system metrics under matched scenario seeds.

The 2000 paired scenarios were generated using a lightweight in-house Python-based traffic-risk simulator rather than CARLA or SUMO. The simulator implements a discrete maneuver set {brake, yield, balanced, proceed, evade}, scenario-dependent occlusion, density, speed, ambiguity, and VRU-presence factors, and group-wise risk estimation for ego occupants, third-party road users, and VRUs. Pedestrians and cyclists in heterogeneous-crowd scenarios are modeled as mixed VRU agents with sampled initial positions, speeds, occlusion exposure, and crossing intent. Their motion follows rule-based constant-velocity or gap-acceptance trajectories with scenario-dependent uncertainty. Fully interactive social-force or game-theoretic pedestrian responses are not modeled and are therefore listed as a limitation. Table 1 summarizes the scenario-generator configuration and heterogeneous VRU modeling assumptions.

Given our hypothesis that averages obscure ethical failures, we adopted a Tail-Focused Reporting strategy. We focus on the “Top-K” outcomes—the worst scenarios in the dataset.

Let K denote the tail set size. We set K=100. For each group i, we extract the top-K risk values from the set {Ri(n)}n=1N. The distribution of these top-100 values characterizes the extreme outcomes (the “tail”).

Beyond tail visualization, we computed four fairness diagnostics: 1.

Per-Scenario Gini Coefficient (G(n))

The Gini coefficient is widely used in economics to measure inequality. Here, it is adapted to quantify risk inequality across the three road-user groups. For a specific scenario n, the Gini coefficient is calculated from the risk vector R(n):(14)G(n)=∑a=1m∑b=1m|Ra(n)−Rb(n)|2m∑a=1mRa(n)+εG,m=3where m=3 is the number of groups, and εG>0 is a small stabilizer. Interpretation: G(n)=0 implies perfect equality (all groups face identical risk). G(n)≈1 implies maximal inequality (one group bears all the risk). A lower mean Gini indicates a method that consistently distributes burden equitably. 2.

Worst-Off Risk (Rmax(n))

This metric tracks the magnitude of risk faced by the least fortunate group in a scenario:(15)Rmax(n)=maxi∈{Ego,Third,VRU}Ri(n)

Reporting the distribution of Rmax(n)—including its mean and tail statistics—enables evaluation of adherence to the Rawlsian maximin principle, which prioritizes minimizing the risk experienced by the worst-off group. 3.

Worst-Off Group Distribution (g(n))

We identify which group is the worst-off in each scenario:(16)g(n)=arg⁡maxi∈{Ego,Third,VRU}Ri(n)

We report the empirical frequencies of g(n) across the 2000 scenarios.

Significance: If g(n)=VRU in a disproportionately high percentage of cases (e.g., >33%), it suggests a systemic bias against vulnerable users. Reducing this frequency is a key ethical objective. 4.

VRU/Ego Tail Ratios

To quantify the specific relationship between VRU and Ego risks in the tail, we compute quantile ratios:(17)Ratiop=Qp(RVRU(n))Qp(REgo(n))+εR,p∈{0.95,0.99}and the Top-K Median Ratio:(18)RatioTopK=median(TopK{RVRU(n)})median(TopK{REgo(n)})+εRwhere εR>0 is a numerical stabilizer. Ratios below unity (Ratio<1.0) indicate that VRU tail risk does not exceed Ego tail risk, consistent with stronger VRU protection.

To avoid ambiguity, this study separates the planner-objective evaluation track from the communication-aware IoV evaluation track. The planner-objective track reports physical collision rate, ex post total harm, Gini coefficient, worst-off risk, and VRU tail risk. The IoV track reports communication-conditioned collision proxy, normalized total harm, latency, AoI, dropout index, and occluded-scene VRU risk. These metrics are used for within-track comparison only and should not be numerically compared across tracks. Table 2 provides compact definitions of the collision-related and harm-related metrics used in the planner-objective and IoV evaluation tracks.

Fig. 1 visualizes the distribution of extreme risks using the top-100 risk values per group. Relative to the baselines, Ethical-Improved compresses the upper tail of VRU risk, consistent with the reductions in VRU p₉₅/p₉₉ in Table 3. Importantly, Ego tail risk does not increase; Ego p₉₅/p₉₉ also decreases to 0.702/0.793, indicating that the tail-regularized objective improves safety without trading VRU protection for greater ego exposure.

For the planner-objective results in Tables 3–8, the Ethical-Improved objective used a static VRU tail threshold calibrated at τ = q_0.95. Communication-state-dependent threshold tightening is reserved for the communication-aware extension and policy-level calibration; the robustness of q_0.90–q_0.99 choices is examined in Section 4.8.

Table 3 reports aggregate safety and efficiency metrics. Ethical-Improved attains the lowest collision rate (0.175), improving on Standard (0.204), Selfish (0.262), and Ethical-Orig (0.221), and also yields the lowest total harm (509.16 vs. 668.98, 988.74, and 700.88, respectively). These gains coincide with reduced tail risks for both Ego and VRU; notably, VRU tail risk declines to 0.591 (p₉₅) and 0.627 (p₉₉). Fig. 2 shows the cumulative personal harm across scenarios for each road-user group.

Beyond aggregate safety, we evaluate distributive fairness using inequality and worst-off diagnostics that explicitly capture ethically salient tail outcomes. 1.

Gini Coefficient (Inequality)

Table 4 summarizes the fairness metrics. Ethical-Improved achieves the lowest mean Gini (0.071) and the lowest 95th-percentile Gini (0.147), indicating a more even allocation of risk across groups than Standard (0.083) and Ethical-Orig (0.095). It also reduces the mean worst-off risk to 0.567, suggesting a systematic reduction in severe outcomes.

The compression of the VRU upper tail in Fig. 1 indicates that the proposed objective does not merely shift the median risk distribution but specifically targets rare high-risk VRU outcomes. This pattern is consistent with the squared hinge tail penalty, which becomes active only when VRU risk exceeds the calibrated threshold. The simultaneous reduction in Ego p₉₅/p₉₉ risk suggests that the improvement is not achieved by transferring tail risk from VRUs to ego occupants, but by encouraging earlier conservative maneuvers in high-uncertainty conflict zones.

Fig. 2 should be interpreted as an aggregate harm decomposition rather than a fairness metric alone. The lower cumulative harm across all groups indicates that the proposed regularizers suppress severe conflict outcomes that contribute disproportionately to total harm. This supports the interpretation that fairness regularization improves safety in the present scenario bank because tail events are also high-harm events.

Fig. 3 supports these results at the distribution level: the per-scenario Gini distribution shifts left under Ethical-Improved (Fig. 3a), and the worst-off risk distribution also shifts downward (Fig. 3b). The Selfish baseline has a modest mean-Gini value but much worse collision and harm results, showing that inequality metrics must be read together with safety and tail diagnostics. 2.

Worst-Off Group Frequencies

Table 5 reports the worst-off-group distribution and VRU/Ego tail ratios. Ethical-Improved reduces the probability that VRUs are the worst-off group to 0.059, compared with 0.319 for Standard, 0.316 for Selfish, and 0.210 for Ethical-Orig. Concurrently, the VRU/Ego tail ratios decline to 0.841 (p₉₅) and 0.791 (p₉₉), reflecting stronger relative protection for VRUs in extreme cases.

The lower-left position of Ethical-Improved in Fig. 4 shows that, in this simulator, fairness and collision reduction are not in conflict. However, this point should not be interpreted as a Pareto-free improvement because Tables 3 and 6 show a clear travel-time cost. The figure therefore supports a constrained policy interpretation: the planner improves safety and distributive fairness, but mobility managers must decide whether the associated delay is acceptable for a given urban context.

Fig. 5 provides a complementary view of worst-off-group frequencies. The stacked distribution shows that Ethical-Improved nearly eliminates VRUs as the maximum-risk group across scenarios. In contrast, the increased share of Ego as the worst-off group is accompanied by a lower overall worst-off severity (Table 4). 3.

Tail Ratios (VRU vs. Ego)

Fig. 6 summarizes VRU-to-Ego tail-risk ratios at p₉₅, p₉₉, and the top-100 median. Ethical-Improved achieves ratios below 1 across all three summaries (0.841, 0.791, and 0.817), indicating that even in the tail the VRU risk remains below the Ego risk.

Fig. 5 indicates a redistribution of worst-off status away from VRUs. This is ethically desirable only when interpreted together with absolute risk levels: the increased Ego worst-off frequency does not imply higher Ego tail exposure, because Ego p₉₅/p₉₉ and mean max risk also decrease. Thus, the figure supports the claim that VRU tail dominance is suppressed without creating an ego-sacrifice policy.

The fact that all VRU/Ego tail ratios remain below 1 under Ethical-Improved indicates that the ratio cap affects the most ethically sensitive part of the distribution rather than only mean outcomes. This provides direct evidence for the protective-ratio mechanism in Eq. (7).

Fig. 4 depicts the fairness-safety trade-off by plotting the mean Gini against collision rate. Ethical-Improved lies in the lower-left region, achieving both lower inequality (Table 4) and lower collision rate (Table 3) than the baselines. This result indicates that fairness regularization can be compatible with safety improvement, although it comes with a measurable efficiency cost.

Risk Redistribution and Efficiency Cost

Ethical-Improved improves safety and tail-risk allocation, but it also imposes a measurable efficiency cost. Average travel time increases from 42.395 s under Standard to 49.481 s under Ethical-Improved, corresponding to a 16.7% increase relative to Standard. Relative to Ethical-Orig, travel time increases from 42.528 to 49.481 s, corresponding to a 16.4% increase. Therefore, the efficiency cost should not be interpreted as negligible; it reflects the operational cost of earlier yielding, more conservative conflict handling, and stronger VRU tail-risk protection in high-risk urban scenarios.

Table 7 reports paired statistical tests over the N = 2000 matched scenarios. For transparency, the p-values shown are the raw paired-test values for the prespecified comparisons, together with mean differences and 95% confidence intervals. Because 11 comparisons are reported in Table 7, Holm familywise correction was additionally applied; all listed comparisons remain significant at α = 0.05 after correction. The statistical evidence should nevertheless be interpreted together with effect sizes and directional consistency.

Table 8 summarizes the paired robustness checks, effect sizes, and confidence intervals for the primary comparison between Ethical-Improved and Ethical-Orig. In addition to p values and confidence intervals, effect magnitudes are reported to avoid interpreting statistical significance alone. Binary collision outcomes are evaluated with McNemar’s exact test and paired risk difference. Continuous paired outcomes are evaluated with permutation tests and Wilcoxon signed-rank robustness checks. For continuous paired outcomes, standardized effect size was computed as Cohen’s dz, defined as dz=D¯sD, where Dn=Xn(Ethical−Improved)−Xn(Ethical−Orig) for matched scenario n. Bootstrap 95% confidence intervals were computed over paired scenario indices. For binary collision outcomes, Cohen’s d is not directly applicable; therefore, McNemar’s paired risk difference and Cohen’s h are reported.

This subsection introduces a second evaluation track that is distinct from the planner-objective comparisons reported in Tables 3–8. Here, the communication-aware objective is evaluated under five connectivity/perception settings—local-only, V2X, RSU-assisted, edge-assisted, and communication-impaired—using common scenario seeds so that the reported differences reflect IoV information quality rather than scenario mismatch. Accordingly, Table 9 summarizes the within-track connected-baseline results, Table 10 documents the communication-state sampling rules, and Table 11 reports the corresponding communication and occlusion metrics. These IoV-track tables should not be directly compared numerically with the planner-baseline results in Tables 3–8.

Table 9 summarizes the IoV-system results, Table 10 documents the communication-state sampling rules, and Table 11 reports the connectivity-baseline communication and occlusion metrics. Edge-assisted cooperative perception achieves the strongest overall trade-off among the connected baselines, yielding the lowest collision proxy (0.225), normalized total harm (291.391), mean Gini (0.046), and VRU p₉₅ risk (0.684). Although edge assistance adds nonzero latency, it improves dropout robustness, lowers the occluded-scenario collision proxy to 0.244, and reduces occluded-scene VRU p₉₅ risk to 0.726.

The communication-state sampling rules are reported in Table 10, and the connectivity-baseline communication and occlusion metrics are reported in Table 11. “Comm.-impaired” denotes communication-impaired operation under degraded V2X/RSU conditions.

The communication-impaired baseline, denoted as “Comm.-impaired” in Tables 9–11, is not a separate planning objective. It uses the same communication-aware Ethical-Improved objective as the other IoV settings but samples z_t from a degraded V2X/RSU operating regime. This regime is characterized by low V2X availability, reduced PDR, high AoI, partial RSU outage, reduced perception confidence, increased edge delay, and higher transmission latency.

The high latency of the communication-impaired condition results from ttx=84ms, edge-delay mean of 28 ms, high AoI, and unstable RSU coverage. Consequently, “Comm.-impaired” represents stale, lossy, and partially unavailable cooperative perception rather than a different ethical planner.

To evaluate parameter robustness, we conducted a compact one-at-a-time sensitivity analysis for the Ethical-Improved planner under the edge-assisted cooperative perception setting. All runs used the same scenario bank, communication-state samples, estimated-risk noise, and collision random numbers, so that differences reflect parameter perturbations rather than random resampling. The tail-risk threshold τ was calibrated from the empirical VRU-risk distribution of the reference yielding maneuver, yielding τ values of 0.639, 0.681, 0.720, and 0.776 for q_0.90, q_0.95, q_0.975, and q_0.99, respectively.

The sensitivity metrics follow the communication-aware IoV track definitions. Table 12 is based on an independent sensitivity-run reference case with common random numbers; therefore, its baseline values are close to, but not identical to, the Edge row in Table 9.

Table 12 reports a compact range summary. Across the tested perturbations, collision proxy remains within 0.222–0.225, mean Gini remains within 0.047–0.054, and average travel time remains within 50.917–51.875 s. The τ sweep shows the expected safety-efficiency pattern: stricter thresholds reduce VRU p₉₅/p₉₉ risk but slightly increase travel time. Fig. 7 visualizes this τ-threshold sensitivity for VRU p95 risk and average travel time. The ρ sweep provides the clearest ratio-control effect, with ρ = 0.8 yielding the lowest VRU/Ego p₉₅ ratio at the cost of higher travel time.

The increase in Ego worst-off frequency under Ethical-Improved requires explicit ethical interpretation. Ethical-Improved increases the frequency with which Ego is the maximum-risk group to 0.539, while reducing the VRU worst-off frequency to 0.059. This does not mean that the proposed planner requires ego occupants to self-sacrifice for VRUs. Rather, it indicates that after disproportionate VRU tail dominance is constrained, the Ego more often becomes the largest remaining risk component in the group-wise ranking. This ranking change should be interpreted together with absolute risk levels: Ego p₉₅/p₉₉ risks decrease from 0.772/0.887 under Ethical-Orig to 0.702/0.793 under Ethical-Improved, and mean max risk decreases from 0.638 to 0.567. Thus, the result reflects a change in relative risk ranking rather than an increase in absolute Ego tail exposure.

Nevertheless, a VRU-protective planner should not be interpreted as authorizing unlimited risk transfer to ego occupants. Practical deployment should include an explicit ego safety floor, hard collision-avoidance constraints, maximum allowable ego-risk thresholds, and legal duty-of-care requirements toward passengers. From a public-acceptance perspective, limited altruistic yielding may be acceptable when it reduces severe VRU harm, but users are unlikely to accept a system that substantially increases passenger danger. The proposed method should therefore be understood as a policy-calibrated ethical planner rather than a universal moral rule.

The results suggest that Ethical-Improved improves both safety and fairness by suppressing severe tail events rather than merely redistributing risk. Compared with Ethical-Orig and the baseline ethical planner in [1], the proposed formulation changes three mechanisms: responsibility is internalized as a bounded risk weight, fairness terms are normalized to reduce scale sensitivity, and VRU tail-risk dominance is penalized directly. The observed reductions in both VRU and Ego p₉₅/p₉₉ risk are consistent with earlier conservative conflict handling, such as braking or yielding before the vehicle enters a high-uncertainty conflict zone.

The 16.7% travel-time increase has practical implications for deployment. At the individual-vehicle level, an additional 7.086 s per high-risk segment may be acceptable in school zones, unsignalized crossings, or dense VRU areas, where severe VRU harm reduction is a priority. At the traffic-system level, however, repeated conservative yielding could reduce intersection throughput, increase queue formation, and create secondary delays during peak demand. User acceptability may therefore depend on whether the delay is transparent, predictable, and limited to high-risk contexts. Mobility managers may need to calibrate the tail-risk threshold and ratio penalty by road type, time of day, VRU density, and congestion state rather than deploy fixed ethical weights across all contexts. The present study does not quantify network-level throughput or queue spillback, so this operational impact should be evaluated in future microscopic and macroscopic traffic simulations before deployment.

No real-world dataset or hardware-in-the-loop experiment was used in the present study. The current validation is limited to controlled paired simulations generated by an in-house lightweight traffic-risk simulator. Therefore, the results should be interpreted as objective-level and mechanism-level evidence rather than deployment-ready validation.

These limitations are prioritized according to their impact on deployment validity. The most important limitation is that the results are obtained from an in-house lightweight simulator; therefore, the conclusions should be interpreted as objective-level and mechanism-level evidence rather than deployment-ready validation. The second limitation is the rule-based VRU model, which does not fully capture interactive pedestrian or cyclist responses. The third limitation is the lightweight communication model, which should be replaced or complemented by packet-level network co-simulation before deployment claims are made. To reduce the limitation of data availability, the generated scenario seeds, aggregate risk logs, metric tables, and analysis scripts will be released through the GitHub repository specified in the Data Availability statement. Table 13 summarizes the prioritized limitations and corresponding deployment-validation priorities.

The proposed framework is locally scalable at the ego-planner level because the optimization is performed over a fixed, discrete set of maneuvers and a small number of risk groups, rather than over all individual agents in the network. The edge/RSU layer supplies communication-conditioned perception inputs, while the ego vehicle solves the risk-allocation objective at each planning cycle. However, large-scale IoV deployment with dense traffic and multiple RSUs was not evaluated in this study. Such deployment would require runtime profiling, multi-RSU coordination, packet-level V2X co-simulation, and stress testing under high vehicle and VRU density. Table 14 summarizes the validation and scalability issues that must be addressed before deployment-ready claims can be made.