MPrISM: Instantaneous Safety Metric for ADS

• MPrISM is an optimization-driven safety metric that quantifies the minimal time before collision using worst-case, best-response strategies.
• It employs minimax quadratic programming over linearized vehicle and pedestrian dynamics to generate theoretically defensible safety certificates in multi-agent scenarios.
• Empirical evaluations via simulations show that MPrISM outperforms traditional metrics like TTC and PCM in terms of ROC AUC and recall, guiding more effective alarm mechanisms.

The Model Predictive Instantaneous Safety Metric (MPrISM) is a formal, optimization-driven real-time safety metric targeted at evaluating the near-term operational safety status of automated driving systems (ADS). It quantifies, for a given traffic snapshot, the minimal future time before a collision becomes unavoidable under worst-case adversarial behavior from one principal traffic agent and best-response evasive action from the subject vehicle. MPrISM employs minimax quadratic programming over linearized vehicle or pedestrian dynamics, yielding theoretically defensible certificates of safety and supporting tractable real-time implementation in large-scale, multi-agent traffic scenarios. Its effectiveness and limitations have been assessed via simulation and systematic performance evaluation frameworks anchored in logged vehicle trajectories (Weng et al., 2020, Yan et al., 2024).

1. Formal Definition and Conceptual Framework

MPrISM defines safety in terms of the shortest time horizon $N\Delta$ (where $\Delta$ is the sensing-action interval) before an unavoidable collision, based on dynamic interaction between the subject vehicle (SV) and its principal adversary (background vehicle, BV). For the SV indexed by $0$ and principal agents by $i=1,\ldots,k$, the joint state vector is $x(t) = [x_0(t), x_1(t), \ldots, x_k(t)]$, with each agent $i$ governed by continuous or discretized motion models (e.g., differential-drive, pedestrian Dubins-like).

The collision set is given by:

$$\Omega_C = \{ x \mid \exists i \in \{1 \ldots k\} : \|x_i - x_0\|_2 \leq C \}$$

where $C$ is the collision radius or distance threshold. For a given traffic snapshot and finite look-ahead $T$, the key function $h^*_i(x(t), t, n, \Delta)$ computes the worst-case, best-response distance for SV and principal agent $i$ at time $t + n\Delta$:

$$h^*_i(x(t), t, n, \Delta) = \min_{u_i(\cdot)} \max_{u_0(\cdot)} \| x_i(t+n\Delta; u_i(\cdot)) - x_0(t+n\Delta; u_0(\cdot)) \|_2$$

Safety is certified over $[t, t + T\Delta]$ if $h^*_i(x(t), t, n, \Delta) \geq C$ for all $i$ and all $n = 1,\ldots,T$. The MPrISM time-to-collision is defined as the smallest $n\Delta$ for which $h^*_i(\cdot, n, \Delta) \leq C$ for any $i$:

$$\tau(x(t)) = \min_{i=1}^k \{ \min \{ n\Delta \mid h^*_i(\cdot, n, \Delta) \leq C \} \}$$

If no such $n$ exists, $\tau = T\Delta$ by convention.

2. Vehicle and Agent Motion Models

Agents in MPrISM may follow either continuous-time or discrete-time dynamics, suitably linearized for tractable prediction. The standard car-like agent kinematics are: $$\dot p = v \cos \phi,\quad \dot q = v \sin \phi,\quad \dot v = a_x,\quad \dot \phi = \frac{a_y}{v}$$ Discretization and linearization about nominal speed $\tilde v$ yields: $$x^+ = A_v x + B_v u, \quad x = [p, q, v, \phi]^T, \quad u = [a_x, a_y]^T$$ Pedestrian agents may use a Dubins-like simplified model with controls on speed and heading rate, similarly linearized. Control admissibility for all agents is restricted to polytopic approximations of friction-limited polygons (e.g., dodecagon or Kamm's circle), expressed as $L_i \bar u_i \leq b_i$ for agent $i$ on the stacked control vector $\bar u_i$.

3. Mathematical Formulation: Minimax Quadratic Program

The minimax collision avoidance is cast as a quadratic game over finite horizon $T$: $$\min_{\bar u_i} \max_{\bar u_0} J(\bar u_i, \bar u_0)$$ subject to: $L_i \bar u_i \leq b_i,\quad L_0 \bar u_0 \leq b_0$ where $J(\bar u_i, \bar u_0)$ is a quadratic function derived from linearized dynamics rollouts,

$$J(\bar u_i, \bar u_0) = \bar u_i^T P \bar u_i + \bar u_0^T Q \bar u_0 + \bar u_i^T R \bar u_0 + U^T \bar u_i + V^T \bar u_0 + H$$

with $P \succ 0,\ Q \succ 0$ representing cost terms for SV and principal adversary, and $R$ the coupling term. This convex–concave quadratic program allows solution via saddle-point methods or direct reformulation to a single QP for positive definite cases. Such structure enables real-time evaluation suitable for ADS deployment.

4. Real-Time Algorithmic Implementation

At each snapshot, the MPrISM algorithm proceeds as follows:

1. Acquire current SV state $x_0(0)$ and each BV state $x_i(0)$.
2. For each principal agent $i$:
   • Construct time-indexed linear dynamical models $A_0(t), B_0(t)$ and $A_i(t), B_i(t)$.
   • Solve the minimax QP or the bilevel QCQP for increasing $N$ until $\|x_i(N) - x_0(N)\|_2 \leq C$.
   • Record $\mathrm{MPrTTC}_i = N\Delta$.
3. Determine overall metric: $\mathrm{MPrTTC} = \min_i \mathrm{MPrTTC}_i$.
4. Compare against alarm threshold $T_{\text{alarm}}$: raise alarm if $\mathrm{MPrTTC} \leq T_{\text{alarm}}$.
5. Repeat at each update interval $\Delta$.

System implementation leverages state-of-the-art QP/QCQP solvers (e.g., Gurobi), warm-starting, and structure-exploiting linear algebra (block-banded matrices) for millisecond-scale inference durations per agent, supporting update rates $>20$ Hz in multi-agent scenes (Weng et al., 2020, Yan et al., 2024).

5. Theoretical Safety Guarantees

MPrISM yields rigorous safety certificates under standard assumptions:

• If $h^*_i(\cdot, n, \Delta) \geq C$ for all $i, n \leq T$, then provably no collision with any principal agent can occur in $[t, t+T\Delta]$.
• If $\tau < T\Delta$, this $\tau$ serves as a conservative lower bound for worst-case time-until-collision, given dynamically feasible input sequences.

Proposition: Under single principal adversary and fixed $\Delta$, if the computed $\tau = T\Delta$, a collision can be guaranteed to be avoidable for duration $T\Delta$; if $\tau < T\Delta$, worst-case collision can occur no later than $t+\tau$.

6. Empirical Performance and Comparative Evaluation

Extensive simulation using SUMO-based datasets (5,050 trips, with both collision and non-collision events, sampled at 10 Hz) enables systematic characterization of detection power using ROC curves (true/false positive rates), precision–recall curves, and alarm timing analysis (Yan et al., 2024). Principal findings include:

• MPrISM achieves highest ROC AUC (predictive power) versus PEGASUS Criticality Metric (PCM) and Time-to-Collision (TTC).
• MPrISM recall approaches $0.95$ with moderate precision ($0.7$–$0.8$). PCM trades off slightly reduced recall but higher precision ($0.85$–$0.95$).
• TTC underperforms except with highly conservative thresholds.
• Earlier alarm requirements reduce AUCs but MPrISM maintains robust trade-off.
• Precision improves slightly for alarms measured $0.5$–$1.0$ s before collision-unavoidable moment.

7. Limitations, Failure Modes, and Refinement Strategies

Failure analysis reveals both false positives and negatives, linked to MPrISM's worst-case behavioral and geometric simplifications:

• False alarms may arise when BV is assumed to steer directly into SV lane under point-mass model.
• False negatives occur for simultaneous lane changes where real vehicle geometry (rectangles) leads to collision not captured by point-mass with radius $C$.
• Overly conservative assumptions (worst-case maneuvers without probabilistic weighting) and simplistic collision geometry (single-point threshold) drive most deficiencies.

Suggested refinements include:

• Employ multi-circle or full-polygon vehicle approximations.
• Adaptive calibration of collision threshold $C$ to balance precision/recall.
• Incorporate maneuver likelihood weighting (e.g., penalize extreme lateral jerk).
• Extend adversarial modeling to multi-BV scenarios in dense traffic conditions.

A plausible implication is that increasing geometric and behavioral model fidelity within the MPrISM framework could further optimize real-time safety status assessment for ADS deployment.

---

MPrISM embodies a mathematically rigorous, adversarially robust approach for real-time risk evaluation in automated driving, integrating dynamic modeling, game-theoretic optimization, and empirical validation. Ongoing research continues to refine its geometric and behavioral assumptions to support operational deployment in complex real-world environments (Weng et al., 2020, Yan et al., 2024).