P-CRI: Prediction-based Collision Risk Index

• Prediction-based Collision Risk Index (P-CRI) defines collision risk measures by mapping predicted future motion into scalar or structured risk quantities over a finite horizon.
• P-CRI employs multiple forms—calibrated probabilities, heuristic surrogates, and routing scores—tailored for specific applications in planning, warning, and control.
• Integrating P-CRI into decision-making processes enhances safety by enabling risk-aware planning and real-time calibration of predictive models.

Prediction-based Collision Risk Index (P-CRI) denotes a family of collision-risk measures that map predicted future motion of an ego agent and surrounding agents or obstacles to a scalar or structured risk quantity over a finite horizon. Recent work does not use the term uniformly, but the underlying construct recurs as a probability of at least one collision, a chance-constraint violation probability, a learned surrogate of collision likelihood, a path-level risk bound, or a routing score for allocating prediction effort (Weiss et al., 24 Jul 2025, Sharma et al., 10 Aug 2025, Wang et al., 2023, Blake et al., 2018, Yang et al., 16 Jul 2025). In this sense, P-CRI is best understood not as a single formula but as a design pattern: define a predictive model, define a collision event, aggregate risk over space and time, and expose the result as a decision variable for planning, warning, or control.

1. Conceptual scope and representative forms

Several representative instantiations illustrate the breadth of the concept.

Framework	Risk quantity	Primary role
GLR (Weiss et al., 24 Jul 2025)	$$P_{\text{collision}} = 1-\exp\!\left(-\int_0^{T_F}\lambda(t)\,dt\right)$$	Horizon collision probability
MonoMPC (Sharma et al., 10 Aug 2025)	$r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$	Risk-aware MPC cost
Prediction-based stochastic FRS (Wang et al., 2023)	$P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$	Highway collision assessment
HyPRAP (Yang et al., 16 Jul 2025)	$$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$	Predictor routing
CATPlan (Xiong et al., 10 Mar 2025)	$$\mathrm{P}(\text{collision}\mid \hat{\tau}_{\text{plan}})$$	Safety monitoring for end-to-end AD

A common misconception is that P-CRI must always be a calibrated probability. The literature includes both explicitly probabilistic quantities and operational surrogates. GLR, stochastic FRS formulations, and chance-constrained MPPI compute probabilities or probability-like quantities over trajectories or time steps (Weiss et al., 24 Jul 2025, Wang et al., 2023, Serfling et al., 27 May 2026). By contrast, MonoMPC uses an MMD-based surrogate of chance-constraint violation, HyPRAP uses a normalized heuristic index derived from prediction-based approach distances and times, and PCMP uses a time-dependent edge risk inside graph search (Sharma et al., 10 Aug 2025, Yang et al., 16 Jul 2025, Hahn et al., 2020). This suggests that P-CRI is best classified by what it operationalizes—warning, constraint enforcement, candidate ranking, or control allocation—rather than by one mandatory probabilistic form.

2. Mathematical structure

A canonical formulation starts from a deterministic or planned ego trajectory and a predictive distribution over other agents. In GLR, the target quantity is the total probability of at least one collision over $[0,T_F]$, formally approximated through a two-stage procedure: first an instantaneous collision probability $P_{\text{col}}(t)$, then a non-homogeneous Poisson process with hazard

$$\lambda(t)=\frac{P_{\text{col}}(t)}{1-P_{\text{col}}(t)},$$

yielding

$$P_{\text{collision}}=1-\exp\!\left(-\int_0^{T_F}\lambda(t)\,dt\right).$$

This is a horizon-level P-CRI in the strongest sense: a scalar in $[0,1]$ with a direct interpretation as probability of at least one collision (Weiss et al., 24 Jul 2025).

A discrete-time variant appears in highway reachability. After propagating a stochastic forward reachable set, the instantaneous collision probability at step $r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$0 is

$r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$1

and the horizon-level collision probability is

$r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$2

The same structural form appears in earlier prediction-based reachability for highway cut-ins, where collision probability is the probability mass of reachable states overlapping the ego trajectory (Wang et al., 2023, Wang et al., 2022).

A different mathematical route appears in MonoMPC. There the random variable is the minimum obstacle clearance $r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$3, with collision encoded through the constraint $r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$4. The nominal risk is the violation probability

$r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$5

but planning uses an MMD-based empirical surrogate,

$r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$6

computed from samples of the residual $r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$7 (Sharma et al., 10 Aug 2025).

Chance-constrained MPPI provides yet another form. There the per-rollout, per-step collision probability $r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$8 is integrated into both a hard constraint and a soft cost, while the executed-step risk is

$r_{\text{MMD}}^{\text{emp}}(\mathbf{u},\mathbf{o})$9

This yields a P-CRI tied to the action actually executed in closed loop, which is especially useful for calibration analysis (Serfling et al., 27 May 2026).

The older FPR formulation shows that a P-CRI can also be a rigorously computable upper bound on path risk. For a candidate path with swept area $P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$0, FPR defines

$P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$1

an efficiently computable upper bound on collision probability that can be used to rank candidate trajectories (Blake et al., 2018).

3. Predictive and geometric models

The predictive object on which P-CRI is built varies substantially across domains. In autonomous racing, GLR assumes an oriented rectangular ego footprint and a rectangular target footprint whose centroid distribution $P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$2 may come from probabilistic Bézier curves or any trajectory predictor. Instantaneous collision is approximated by evaluating whether five representative target points—the four corners and the centroid—fall inside the ego rectangle, with 2D Gauss–Legendre cubature used for the spatial integrals (Weiss et al., 24 Jul 2025).

MonoMPC replaces explicit vehicle–vehicle overlap with a clearance random variable. A monocular RGB image is processed by a pre-trained monocular depth estimator, projected into a 2D point cloud, encoded by PointNet++, and combined with robot state and candidate control sequence in an MLP that predicts the mean, variance, and kernel width of the minimum-clearance distribution. The risk variable is therefore not overlap probability directly but the distribution of worst-case obstacle clearance along the candidate trajectory (Sharma et al., 10 Aug 2025).

Highway reachability methods adopt a state-space formulation. In the confidence-aware probabilistic collision detection framework, the surrounding vehicle is propagated through a stochastic forward reachable set with input probabilities obtained from an LSTM-based acceleration model and a confidence-aware dynamic belief over scaling parameters $P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$3. In the earlier prediction-based reachability method, the state is a 2D point-mass model with state $P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$4, and multi-modal bivariate acceleration distributions drive the state transition matrix online (Wang et al., 2023, Wang et al., 2022).

Urban surveillance and assistive-navigation work use still different geometry. The crosswalk PCRA method predicts future speed and heading with deep LSTM networks, then constructs predictive collision risk areas as truncated circular sectors from ECDF-derived confidence intervals; severity is divided into danger, warning, and relatively safe according to overlap at about 1, 2, and 3 seconds (Noh et al., 2021). The pedestrian-navigation risk model for people with vision impairment computes collision probability from overlap of Gaussian predictive position densities of the user and nearby objects and integrates those probabilities over a horizon through a survival function (Tourki et al., 19 Jun 2025).

This diversity implies that P-CRI is not tied to any single geometry class. Rectangles, disks, point masses, occupancy fields, clearance variables, and graph edges all appear as the latent state on which risk is computed.

4. Integration into planning, warning, and control

A central use of P-CRI is as a constraint or cost in online decision making. GLR is explicit: the horizon collision probability can be used as a chance constraint,

$P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$5

or as a cost term

$P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$6

with $P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$7 denoting the nominal performance objective (Weiss et al., 24 Jul 2025).

MonoMPC embeds its risk surrogate directly in a receding-horizon objective,

$P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$8

subject to dynamics and control bounds. The risk term is evaluated by Monte Carlo sampling from the learned clearance distribution and by computing the MMD between the residual distribution and a Dirac distribution at zero residual (Sharma et al., 10 Aug 2025).

In chance-constrained MPPI, the same pattern appears in a stochastic-control setting. Per-step collision probabilities $P_{\text{col}} = 1-\prod_k (1-p_{\text{hit}}(k))$9 are added as a soft cost,

$$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$0

and rollouts are rejected if $$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$1 for any $$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$2 (Serfling et al., 27 May 2026).

Graph-based planning gives a more discrete integration. In PCMP, the edge cost is

$$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$3

where $$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$4 is the time-period dependent collision risk on an edge derived from predicted obstacle occupancy. The risk parameter $$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$5 controls the path-length versus risk trade-off (Hahn et al., 2020).

HyPRAP uses P-CRI differently. Its $$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$6 does not directly enter a collision-avoidance cost; instead it routes each obstacle to a high-accuracy predictor, a cheaper predictor, or effectively no predictor, via thresholds $$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$7. The resulting predictions and conformal radii then define the collision-avoidance constraints in MPC (Yang et al., 16 Jul 2025).

RCMS combines instantaneous and predictive risk at the supervisory level. Activation occurs if either overlap-based risk $$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$8 or TTCE-based predictive risk $$\psi(t,k)=w_1 g_1(\text{PAD}(t,k))+w_2 g_2(\text{PAT}(t,k))$$9 exceeds its activation threshold; deactivation requires both to fall below lower thresholds. Once active, RCMS minimizes a horizon cost

$$\mathrm{P}(\text{collision}\mid \hat{\tau}_{\text{plan}})$$0

where $$\mathrm{P}(\text{collision}\mid \hat{\tau}_{\text{plan}})$$1 is a smooth situational risk field built from surrounding vehicles and road boundaries (Tariq et al., 2023).

5. Empirical behavior, calibration, and comparison

Reported results show that predictive collision indices matter primarily through the joint accuracy–latency–calibration trade-off. In 446 overtaking scenarios from a high-fidelity Formula One racing simulation, GLR reduced average error by 77%, surpassed the next-best method by 52%, and ran at 1000 Hz (Weiss et al., 24 Jul 2025). In HyPRAP, P-CRI-based routing achieved 93.1% success with 592 total model calls, whereas a proximity-only baseline at similar computation achieved 80.3%, and a proximity-only baseline tuned to similar success required 1590 total calls (Yang et al., 16 Jul 2025).

Risk-aware learned models also show large practical gains. MonoMPC reported 9x and 7x improvements in success rates over NoMaD and the ROS stack, respectively, and its ablations showed that NLL-only variance learning and near-deterministic variance both degraded safety substantially (Sharma et al., 10 Aug 2025). CATPlan, which predicts $$\mathrm{P}(\text{collision}\mid \hat{\tau}_{\text{plan}})$$2 from internal motion and planning queries of end-to-end planners, improved average precision by 54.8% relative to a GMM-based baseline on NeuroNCAP (Xiong et al., 10 Mar 2025). In pedestrian navigation for persons with vision impairment, the probabilistic risk model reached 67% warning accuracy on real-world data, whereas distance and time-to-contact reached 51% (Tourki et al., 19 Jun 2025).

Calibration emerges as a distinct concern rather than a side issue. Chance-constrained MPPI under dual uncertainty evaluates executed-step risk $$\mathrm{P}(\text{collision}\mid \hat{\tau}_{\text{plan}})$$3 against actual collisions $$\mathrm{P}(\text{collision}\mid \hat{\tau}_{\text{plan}})$$4 using the Brier Score,

$$\mathrm{P}(\text{collision}\mid \hat{\tau}_{\text{plan}})$$5

and Log Loss,

$$\mathrm{P}(\text{collision}\mid \hat{\tau}_{\text{plan}})$$6

That work shows characteristic failure modes: overconfidence causes systematic safety violations, underconfidence induces freezing, and dual underconfidence can produce probability dilution, where inflated uncertainty lowers local collision probabilities and paradoxically yields unsafe aggression (Serfling et al., 27 May 2026). This suggests that a P-CRI is only as reliable as the uncertainty estimates that feed it.

Performance can also be scenario-dependent rather than absolute. PCMP showed that a risk-sensitive agent avoided 47.3% of collision scenarios with a 1.3% detour, while a risk-averse agent avoided up to 97.3% with a 39.1% detour (Hahn et al., 2020). Confidence-aware stochastic FRS reduced false positives and improved timeliness in highway cut-ins relative to heuristic and non-adaptive baselines (Wang et al., 2023). These results reinforce that P-CRI design is inseparable from the operating point at which warnings, interventions, or candidate selections are made.

6. Assumptions, limitations, and extension paths

The main limitations recur across formulations. GLR assumes independence among representative corner and centroid random variables, adopts a heuristic log-odds mapping from instantaneous collision probability to hazard, and treats the ego trajectory as deterministic (Weiss et al., 24 Jul 2025). Reachability-based highway methods assume structured lane-based environments, limited maneuver classes such as keep lane and lane change left or right, and effectively perfect state information for building reachable sets and predictor inputs (Wang et al., 2023, Wang et al., 2022). HyPRAP’s $$\mathrm{P}(\text{collision}\mid \hat{\tau}_{\text{plan}})$$7 is explicitly a heuristic normalized index rather than a calibrated probability (Yang et al., 16 Jul 2025). DUCCT-MPPI assumes Gaussian localization and obstacle-prediction uncertainty, conditional independence across obstacles, and a one-tube Unscented Transform approximation shared across rollouts (Serfling et al., 27 May 2026). CATPlan depends on the upstream planner’s internal representations; when an object is not detected, the risk head has no corresponding motion query, and collision prediction can fail (Xiong et al., 10 Mar 2025).

A second limitation is semantic non-uniformity. Some works output a true horizon probability, some a per-step probability, some a surrogate loss, some an ordinal severity level, and some a routing score. This suggests that the term P-CRI is most coherent when anchored to four questions: what future is predicted, what collision event is defined, how risk is aggregated over space and time, and how the resulting quantity is consumed by the downstream system.

The extension path is correspondingly broad. The literature already shows transfer from autonomous racing to broader motion planning (Weiss et al., 24 Jul 2025), from ADAS-style probabilistic risk models to pedestrian assistive navigation (Tourki et al., 19 Jun 2025), from crosswalk CCTV to city-scale data-cube analysis via PCR levels (Noh et al., 2021), and from single-predictor control to hybrid predictor routing in dense dynamic environments (Yang et al., 16 Jul 2025). A plausible implication is that future P-CRI research will continue to move along two coupled axes: richer predictive models, including calibrated multi-agent uncertainty, and tighter integration with downstream decisions, including chance constraints, safety monitors, and explanation modules.