VRisker: Risk-Aware Diversification & Simulation

• VRisker is a risk-aware framework that minimizes tail risk by optimizing a CVaR-style metric to ensure minority intents are adequately represented in information retrieval.
• In autonomous driving, VRisker provides a configurable, lightweight simulation framework using rule-based FSM and idealized V2X awareness to validate vulnerable road user safety under Euro NCAP protocols.
• Extensions of VRisker explore planner-aligned risk fields and rare-event simulations, demonstrating its adaptability across diverse applications while highlighting current limitations in standardization.

“VRisker” is used in the supplied literature to denote more than one technically distinct system. In information retrieval, it names a fast greedy re-ranker that treats diversification as a risk-minimization problem within a single query by optimizing VRisk, a CVaR-style metric over the least-served fraction of user intents (Takehi et al., 26 Oct 2025). In vulnerable-road-user safety research, the same label is used for a configurable, lightweight simulation framework grounded in Euro NCAP VRU testing protocols, combining a rule-based finite-state machine, ego-vehicle perception, and idealized V2X awareness for repeatable safety validation (He et al., 21 Oct 2025). Related synthesis documents further describe planner-aligned risk fields, VR rare-event simulation, and Vehicle-in-Virtual-Environment workflows as ways to build or extend a system called VRisker, which suggests that the term is not yet standardized across domains (Thalapanane et al., 2024, Link et al., 20 May 2026, Chen et al., 30 Aug 2025).

1. Diversification as tail-risk minimization

In the information-retrieval usage, VRisker is motivated by ambiguous or underspecified queries such as “jaguar,” for which a single query can correspond to multiple mutually exclusive intents with skewed prior probabilities. Naive ranking schemes that optimize a single expected-relevance score align strongly with dominant intents and can ignore minority intents entirely. The central claim of the underlying formulation is that robustness should be defined not only across queries, but also within a single query, by minimizing the severity and frequency of intent-level failures (Takehi et al., 26 Oct 2025).

The formal setup assumes a query $q$ with finite intent set $C(q) = \{c_1,\dots,c_m\}$ and priors $\Pr(c \mid q)$ satisfying $\sum_{c \in C(q)} \Pr(c \mid q)=1$. For a candidate set $D$, per-intent relevance is $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$, and raw expected relevance is

$$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$

For a top-$k$ ranking $R \subseteq D$, the simplest modular base metric used throughout is average relevance,

$$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$

A key observation is that the “standard” intent-marginal value

$C(q) = \{c_1,\dots,c_m\}$0

is identical to the intent-weighted objective

$C(q) = \{c_1,\dots,c_m\}$1

whenever $C(q) = \{c_1,\dots,c_m\}$2 is linear in the ranking counts, including average relevance, DCG, and Precision@$C(q) = \{c_1,\dots,c_m\}$3. This identity is the stated reason why many diversification methods do not protect minority intents under linear bases. Even under non-linear bases such as nDCG and ERR, the supplied account reports a toy counterexample in which NDCG-IA favors a ranking that leaves 49% of users unsatisfied (Takehi et al., 26 Oct 2025).

The corresponding interpretation of VRisker is therefore not “diversification by average coverage,” but diversification by explicit tail protection. The objective is to reduce poor experiences for the worst-served fraction of intents, rather than merely improve an average over intents.

2. VRisk and the optimization problem

VRisk is defined through a per-intent hinge loss against a target level,

$C(q) = \{c_1,\dots,c_m\}$4

where the default target is the per-intent oracle,

$C(q) = \{c_1,\dots,c_m\}$5

The formulation also permits baseline-relative targets, such as the value attained by a reference system (Takehi et al., 26 Oct 2025).

For a pessimism parameter $C(q) = \{c_1,\dots,c_m\}$6, VRisk measures the expected loss over the worst $C(q) = \{c_1,\dots,c_m\}$7-fraction of the intent probability mass. Using the Rockafellar–Uryasev CVaR representation, it is written as

$C(q) = \{c_1,\dots,c_m\}$8

When $C(q) = \{c_1,\dots,c_m\}$9, this reduces to expected loss,

$\Pr(c \mid q)$0

so the metric generalizes intent-weighted averaging while making the tail explicit. As $\Pr(c \mid q)$1 decreases, evaluation becomes more pessimistic and focuses on rarer tail conditions (Takehi et al., 26 Oct 2025).

VRisker solves

$\Pr(c \mid q)$2

The algorithm is a greedy re-ranker that builds the ranking from the top position down. At each step it adds the document that most reduces $\Pr(c \mid q)$3; ties are broken by maximizing $\Pr(c \mid q)$4. The induced risk-reduction set function

$\Pr(c \mid q)$5

is monotone, and for modular base metrics it is monotone submodular. This yields the classical greedy guarantee

$\Pr(c \mid q)$6

The optimization problem is nevertheless NP-hard for variable $\Pr(c \mid q)$7 and any $\Pr(c \mid q)$8, via reduction from Weighted Max-$\Pr(c \mid q)$9-Cover. For non-modular bases such as nDCG, the supplied result gives a submodularity-ratio guarantee with discount weights $\sum_{c \in C(q)} \Pr(c \mid q)=1$0 and lower bound

$\sum_{c \in C(q)} \Pr(c \mid q)=1$1

implying

$\sum_{c \in C(q)} \Pr(c \mid q)=1$2

Typical lower bounds reported for $\sum_{c \in C(q)} \Pr(c \mid q)=1$3 are $\sum_{c \in C(q)} \Pr(c \mid q)=1$4, $\sum_{c \in C(q)} \Pr(c \mid q)=1$5, and $\sum_{c \in C(q)} \Pr(c \mid q)=1$6 for $\sum_{c \in C(q)} \Pr(c \mid q)=1$7, $\sum_{c \in C(q)} \Pr(c \mid q)=1$8, and $\sum_{c \in C(q)} \Pr(c \mid q)=1$9, respectively (Takehi et al., 26 Oct 2025).

3. Empirical characteristics of the retrieval algorithm

The reported empirical evaluation covers three domains: NTCIR INTENT-2 with 95 queries and approximately 6.1 intents per query, TREC Web 2012 with 50 queries and approximately 6.0 intents per query, and MovieLens 32M with 42,902 users treated as queries and 8 genres treated as intents. In MovieLens, multi-genre items use a Bayes-consistent allocation to per-intent relevance so that expected relevance matches observed ratings (Takehi et al., 26 Oct 2025).

Default settings are $D$0, $D$1, and $D$2 equal to the per-intent oracle. Across AvgRel, nDCG, DCG, ERR, RBP, and Precision@$D$3, the main findings are that VRisker reduces worst-case intent failures by 20–40% relative to naive ranking, reaches reductions up to 33% at longer lists such as INTENT-2 at $D$4, and incurs a standard-performance drop of 0–10%, with a best case of only approximately 2% at $D$5 (Takehi et al., 26 Oct 2025).

The supplied account emphasizes tunability through $D$6. Lower $D$7 yields more robustness and lower VRisk, at the cost of average performance; the example given is $D$8, which yields approximately 20–30% less worst-case loss with approximately 3–5% utility drop. As ranking depth grows, VRisker maintains consistent tail-risk reductions while average utility improves. Under Gaussian noise injected into intent priors, all methods degrade, but VRisker degrades more gracefully and remains best on tail risk; at high noise variance $D$9, it can match or exceed baseline average performance (Takehi et al., 26 Oct 2025).

Mixed-integer programs are used to compute optimal VRisk solutions and show that the greedy method is close to optimal across both modular and non-modular bases. Runtime is also a central part of the argument: naïve evaluation is $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$0, but incremental updates reduce the complexity to $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$1, matching the profile of standard greedy diversification methods such as xQuAD, ERR-IA, and IA-Select. On MovieLens 32M, with approximately $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$2 candidates per query, VRisker runs in approximately $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$3 ms/query (Takehi et al., 26 Oct 2025).

Operational guidance in this formulation is explicit. $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$4 is chosen to reflect risk tolerance; $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$5 is chosen to encode a service-level objective; and the required inputs are an intent taxonomy $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$6, intent priors $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$7, and per-intent relevance $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$8. The method is described as agnostic to the candidate source and can be integrated as a post-ranking stage over BM25, a neural re-ranker, or a collaborative filter (Takehi et al., 26 Oct 2025).

4. VRU-safety simulation framework under Euro NCAP protocols

In the vulnerable-road-user safety usage, VRisker is described as a configurable, lightweight simulation framework for assessing and mitigating risk to vulnerable road users through rapid, repeatable validation. It follows Euro NCAP’s AEB/LSS VRU test protocol v4.5.1 and is organized into three pipeline stages: scenario design, environment modeling, and simulation analysis. The architecture combines a rule-based finite-state machine motion planner, ego-vehicle perception through a sector-shaped field of view, and idealized Vehicle-to-Everything awareness (He et al., 21 Oct 2025).

The scenario-definition layer instantiates standardized Euro NCAP VRU tests, including scene layout, conflict geometry, target profiles and paths, and initial states. It exposes ODD parameters such as speed limits, lane topology, and intersection geometry, as well as advanced-feature toggles such as V2X on/off and sensor FOV settings. Dummy VRU targets are used to ensure repeatability. The ego vehicle follows a prescribed path under a kinematic bicycle model, while longitudinal behavior is governed by a discrete-state controller implementing preventive driving and collision avoidance (He et al., 21 Oct 2025).

The framework digitizes three representative Euro NCAP categories that are explicitly illustrated and run in the experiments. The first is Car-to-Pedestrian Nearside Adult, an orthogonal crossing on a straight road. The second is Car-to-Bicyclist Longitudinal Adult, implemented as a same-lane e-scooter-leading case, with the note that a cyclist variant can be set similarly. The third is Car-to-Motorcyclist Front Turn Across Path at an intersection (He et al., 21 Oct 2025).

For the pedestrian case, the initial conditions are specified as ego position $$\mathrm{rel}(d \mid q,c) \in [0,\mathrm{rel}_{\max}]$$9, $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$0, $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$1, $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$2, and pedestrian position $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$3 with $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$4. In the example run, without V2X the system detects at $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$5 s and collides at $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$6 s; with V2X, proactive deceleration occurs at $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$7 s, the pedestrian clears the path by $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$8 s, and no collision occurs (He et al., 21 Oct 2025).

For the e-scooter-leading case, the ego starts at $$\mathrm{rel}(d \mid q)=\sum_{c \in C(q)} \Pr(c \mid q)\,\mathrm{rel}(d \mid q,c).$$9 with $k$0, and the e-scooter starts at $k$1 with $k$2. The trigger condition is relative closing in the same lane, and the goal is acceptable headway rather than merely no collision. In the example run, the no-V2X system detects at $k$3 s and achieves safe headway by $k$4 s; the V2X case triggers earlier and induces more conservative deceleration, but both variants avoid collision (He et al., 21 Oct 2025).

For the motorcyclist FTAP case, the ego starts at $k$5 with $k$6, while the motorcyclist starts at $k$7 with $k$8. Without V2X, the oncoming rider is detected late at $k$9 s and collision occurs at $R \subseteq D$0 s. With V2X, a broadcast arrives at $R \subseteq D$1 s, early braking is triggered, the motorcyclist passes safely, and no collision occurs (He et al., 21 Oct 2025).

5. Motion planning, sensing, and safety indicators

The motion planner in this simulation-oriented VRisker is a rule-based FSM with four states: Cruise, Pre-Slow, Emergency-Brake, and Resume. Cruise maintains desired speed along the path when no VRU hazard is present. Pre-Slow applies gentle preventive deceleration when a VRU is detected beyond a safe-distance threshold. Emergency-Brake applies maximal braking when safe-distance criteria are violated and conflict imminence is detected. Resume transitions back to Cruise when the VRU clears the conflict zone and margins recover (He et al., 21 Oct 2025).

The primary threshold is the two-second rule,

$R \subseteq D$2

with units in meters when $R \subseteq D$3 is in m/s. If VRU distance exceeds $R \subseteq D$4, the controller transitions from Cruise to Pre-Slow; otherwise it escalates from Pre-Slow to Emergency-Brake. Exit conditions are defined in terms of PET, TTC, or minimum-distance recovery. A representative fragment is explicitly given: if perceive_vru() is true and $R \subseteq D$5, then state = PreSlow; otherwise state = EmergencyBrake; if no perception occurs, the controller remains in Cruise (He et al., 21 Oct 2025).

Perception is modeled through a deterministic sector-shaped field of view with configurable range $R \subseteq D$6 and angle $R \subseteq D$7, with experimental defaults of $R \subseteq D$8 m and $R \subseteq D$9. Detection occurs only when the VRU lies inside the FOV sector. Idealized V2X awareness is modeled as a zero-latency, reliable broadcast channel fusing VRU state messages from infrastructure into the perception stack; perceive_vru() returns true if either onboard detection or V2X triggers first. The stated effect is that earlier awareness shifts FSM transitions earlier, increasing TTC and minimum distance in crossing and FTAP conflicts, while improving comfort in longitudinal following by smoothing deceleration (He et al., 21 Oct 2025).

The framework computes several surrogate safety indicators. For a 1D closing scenario,

$$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$0

and

$$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$1

PET is

$$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$2

Stopping distance is

$$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$3

and minimum distance is

$$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$4

Conflict count is defined as the number of events for which $$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$5 or $$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$6 or $$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$7 (He et al., 21 Oct 2025).

Safe/unsafe declarations are collision-based and threshold-based. Any collision sets the safety score to $$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$8 for the scenario. A safe run is collision-free and keeps TTC, PET, and $$V(R \mid q,c)=\frac{1}{k}\sum_{d \in R}\mathrm{rel}(d \mid q,c).$$9 above critical thresholds. A sustainability score from prior work normalizes safety, efficiency, and comfort to $C(q) = \{c_1,\dots,c_m\}$00. In the pedestrian case, the reported scores are safety $C(q) = \{c_1,\dots,c_m\}$01 with V2X versus $C(q) = \{c_1,\dots,c_m\}$02 without, efficiency $C(q) = \{c_1,\dots,c_m\}$03 versus $C(q) = \{c_1,\dots,c_m\}$04, and comfort $C(q) = \{c_1,\dots,c_m\}$05 versus $C(q) = \{c_1,\dots,c_m\}$06. In the e-scooter-leading case, safety is $C(q) = \{c_1,\dots,c_m\}$07 for both, while V2X slightly reduces efficiency and comfort due to conservatism. In the motorcyclist FTAP case, the reported scores are safety $C(q) = \{c_1,\dots,c_m\}$08 with V2X versus $C(q) = \{c_1,\dots,c_m\}$09 without, efficiency $C(q) = \{c_1,\dots,c_m\}$10 versus $C(q) = \{c_1,\dots,c_m\}$11, and comfort $C(q) = \{c_1,\dots,c_m\}$12 versus $C(q) = \{c_1,\dots,c_m\}$13 (He et al., 21 Oct 2025).

6. Related extensions: rare-event VR simulation, planner-aligned risk fields, and VVE bicyclist safety

Related documents in the supplied literature describe how a system called VRisker can be extended beyond Euro NCAP-style rule-based simulation. One line of work adapts TRAVERSE into a VR-based risk and rare-event simulation and assessment framework for autonomous vehicles. That stack combines SUMO for microscopic traffic, Unity for rendering and VR interfacing, and ROS/ROS2 compatibility via ROS-TCP. It uses portable hardware consisting of a Meta Quest Pro and Logitech G29 steering wheel and pedals, and implements NHTSA-inspired pre-crash scenarios including sudden lane change, T-bone crash, sudden vehicle stop in front, vehicle running red lights, sudden deer crossing, crash at roundabout, crash in ramp merger, and jaywalking pedestrian crash. A user study with $C(q) = \{c_1,\dots,c_m\}$14 participants compared TRAVERSE with DReyeVR over approximately five minutes each in randomized order and reported lower overall simulator sickness and significant improvements in sense of presence, ease of adjustment, scenario realism, control responsiveness, audio immersiveness, head tracking, realistic control, and overall experience, while traffic-simulation ratings were not significantly different (Thalapanane et al., 2024).

A second extension uses MC-Risk as a planner-aligned risk-estimation backbone. MC-Risk constructs a bird’s-eye-view risk field by linearly composing a motorized-agent field, a VRU risk field, and a road penalty field. The motorized-agent field fuses a black-box multimodal trajectory predictor with an analytic Gaussian-torus construction; the VRU field uses a forward-biased anisotropic kernel aligned to heading and speed; and the road penalty field exploits HD-map topology with off-road, same-direction, and opposite-direction lane penalties. The field is directly consumable by MPC through an objective of the form

$C(q) = \{c_1,\dots,c_m\}$15

On RiskBench’s collision subset, MC-Risk reports $C(q) = \{c_1,\dots,c_m\}$16, $C(q) = \{c_1,\dots,c_m\}$17, $C(q) = \{c_1,\dots,c_m\}$18, $C(q) = \{c_1,\dots,c_m\}$19, $C(q) = \{c_1,\dots,c_m\}$20, and $C(q) = \{c_1,\dots,c_m\}$21, with the largest ablation degradation arising from removal of the forward-biased VRU component (Link et al., 20 May 2026).

A third extension is VVE-based bicyclist-safety evaluation. In that formulation, VRisker is built around a Vehicle-in-Virtual-Environment method in which a real vehicle drives in a safe space while its state is synchronized to a virtual ego vehicle in CARLA/Unreal. The architecture includes RTK GPS with dual antennas, a dSPACE MicroAutoBox ECU, CAN actuation, Ethernet UDP links to the virtual world, and optional BLE-based V2P/VRU localization. The decision stack is hierarchical: a DQN or DDQN high-level policy outputs discrete actions, while a CLF–CBF–QP or HOCLF–HOCBF–QP controller computes continuous steering and braking subject to barrier constraints such as

$C(q) = \{c_1,\dots,c_m\}$22

and

$C(q) = \{c_1,\dots,c_m\}$23

The study reports low-level control at $C(q) = \{c_1,\dots,c_m\}$24 Hz with approximately $C(q) = \{c_1,\dots,c_m\}$25 ms per step, high-level policy at $C(q) = \{c_1,\dots,c_m\}$26 Hz with approximately $C(q) = \{c_1,\dots,c_m\}$27 ms per step, and a representative injected delay of $C(q) = \{c_1,\dots,c_m\}$28 s. It reproduces five FARS bicyclist crash types—230, 310, 145, 220, and 210—and reports that minimum distance in dynamic interaction tests remained above the “socially acceptable” threshold of $C(q) = \{c_1,\dots,c_m\}$29 m when the CBF safety filter was active (Chen et al., 30 Aug 2025).

Taken together, these related sources show a family of uses in which VRisker denotes either a concrete simulator or a broader research program centered on repeatable hazard generation, risk scoring, and planner-facing safety evaluation for autonomous driving and VRUs.

7. Limitations, assumptions, and unresolved standardization

The information-retrieval VRisker assumes mutually exclusive intents, known or estimated intent priors, and known or estimated per-intent relevance. Its main guarantees depend on modular base metrics, while the non-modular case uses a weaker submodularity-ratio argument. The formulation is single-query and non-personalized at the intent level, and the supplied discussion identifies extensions to overlapping or hierarchical intents, session-level risk objectives, and user-personalized $C(q) = \{c_1,\dots,c_m\}$30 or $C(q) = \{c_1,\dots,c_m\}$31 as future directions (Takehi et al., 26 Oct 2025).

The Euro NCAP-oriented VRU simulator assumes deterministic detection, no sensor noise, no latency, and idealized zero-latency reliable V2X. VRUs are modeled as point-mass kinematic agents on prescribed paths, the planner handles preventive deceleration and emergency braking but not lateral evasive maneuvers or predictive intent inference, and environmental conditions are held fixed for repeatability. The supplied account explicitly notes the resulting simulation-to-real gap and the lack of real-world complexities such as multi-agent interactions, infrastructure variability, and human intent unpredictability (He et al., 21 Oct 2025).

The related extensions preserve additional limitations. TRAVERSE reports that traffic realism ratings were not significantly different from the baseline and notes medium fidelity under Wynne et al.’s criteria (Thalapanane et al., 2024). MC-Risk demonstrates a plug-and-play MPC interface but does not report closed-loop planning results, and its road penalty field depends on accurate HD-map topology (Link et al., 20 May 2026). The VVE bicyclist-safety work uses representative fixed delays, simplified vehicle models, and does not yet deploy all emergency-braking functions on a real vehicle in VVE (Chen et al., 30 Aug 2025).

A plausible implication is that “VRisker” currently functions less as a single canonical system than as a cross-domain label for methods that make risk explicit: in retrieval, by minimizing CVaR-style tail loss over intents; in autonomous driving, by turning rare, safety-critical VRU interactions into repeatable, measurable validation problems; and in planner-aligned risk estimation, by converting scene structure into control-relevant risk fields. The supplied literature therefore supports a conceptually unified theme—risk-aware treatment of minority, tail, or hazardous cases—even though the implementations, assumptions, and evaluation protocols remain domain-specific.