Ethics-Aware Safe Reinforcement Learning for Rare-Event Risk Control in Interactive Urban Driving

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In 2022, 7,522 pedestrians were killed and more than 67,000 injured in the U.S. according to NHTSA.

“According to the U.S. National Highway Traffic Safety Administration, 7,522 pedestrians were killed and more than 67,000 injured in 2022 [3].”

• cites [3] sourcePartly supported

  The Introduction reports 7,522 U.S. pedestrian deaths in 2022, and its cited source is the National Highway Traffic Safety Administration’s Fatality Analysis Reporting System (References, item 1). Our copy does not report more than 67,000 pedestrian injuries; searches for that figure and for pedestrian injuries did not find support. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

Vulnerable road users such as pedestrians and cyclists bear a disproportionately high burden of severe injury and death in collisions.

“This question is especially urgent for vulnerable road users (VRUs) such as pedestrians and cyclists, who bear a disproportionately high burden of severe injury and death in collisions.”

• cites [3] sourcePartly supported

  The Introduction (page unavailable) reports a rapid rise in U.S. pedestrian deaths, and the ‘What is added by this report?’ section reports higher U.S. pedestrian death rates than in the other countries studied. The paper does not establish that pedestrians and cyclists together bear a disproportionate burden of severe injury and death relative to other road users; it does not report cyclist deaths or severe-injury rates. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

A quartic polynomial in the longitudinal direction and a quintic polynomial in the lateral direction guarantee continuity of both longitudinal and lateral accelerations, contributing to a smooth and comfortable trajectory.

“To ensure the comfort of the generated trajectory ξ e,t at time step t , a quartic polynomial in the longitudinal direction l and a quintic polynomial in the lateral direction d are employed, such that, ξ e,t = { l e,τ , d e,τ } . This formulation follows established practice, as it guarantees the continuity of both longitudinal and lateral accelerations, thereby contributing to a smooth and comf”

• cites [30] sourceCould not verify

  The Method (p. 3) describes a GAN-based pedestrian trajectory predictor, while the Conclusion (p. 8) discusses socially acceptable predictions. Searches for quartic and quintic polynomials, longitudinal and lateral motion, and acceleration continuity found no passage addressing the claimed formulation or its comfort guarantee. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

The Stanley controller is particularly effective because it accounts for both lateral displacement and heading errors.

“It is particularly effective because it accounts for both lateral displacement and heading errors.”

• cites [31] sourceCould not verify

  The Metrics and Results sections discuss pedestrian trajectory forecasting, not the Stanley controller. Searches for the Stanley controller, lateral displacement, heading error, and vehicle path tracking found no passage addressing the claimed reason for its effectiveness. Page numbers were not available in our copy. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Ethical preferences vary widely across cultures, as demonstrated by the Moral Machine Experiment.

“Empirical work such as the Moral Machine Experiment has demonstrated that ethical preferences vary widely across cultures [8].”

• cites [8] sourceSupported

May combine two claims.

Under the assumptions that terminal lateral acceleration and jerk are zero, the resulting trajectory minimizes jerk, thereby enhancing passenger comfort.

“Under these assumptions, the resulting trajectory minimizes jerk, thereby enhancing passenger comfort [30].”

• cites [30] sourceCould not verify

  The Introduction (p. 1) and Conclusion (p. 8) describe predicting socially acceptable pedestrian trajectories, not constructing minimum-jerk trajectories. Searches for terminal lateral acceleration, jerk minimization, and passenger comfort found no passage addressing those claims. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

• AssumptionThe authors state this as an assumption: The terminal lateral acceleration and jerk are zero (¨ l T = 0, ˙ d T = 0, and ¨ d T = 0).

May combine two claims.

For the EthicAR agent, the Lagrange multiplier λ initially rises, then oscillates, and eventually plateaus at levels that maintain constraint satisfaction without excessive penalization.

“For the EthicAR agent, λ initially rises as the agent explores and frequently violates cost constraints, thereby enforcing constraint adherence. As training progresses, λ oscillates to balance rewards and costs before eventually plateauing at levels that maintain constraint satisfaction without excessive penalization.”

Classified as probably not a substantial claim. May combine two claims.

Although ego risk values remain low and vary little between agents, the differences are still informative.

“Although these risk values remain low and vary little between agents, the differences are still informative.”

May combine two claims.

The approach outperforms baseline methods in reducing risk to others while maintaining ego performance and comfort.

“demonstrate that it outperforms baseline methods in reducing risk to others while maintaining ego performance and comfort.”

May combine two claims.

EthicAR in selfish mode shows mildly discomforting jerk actions, though these remain within the acceptable range.

“When evaluating jerk, the selfish mode of EthicAR again shows some mildly discomforting actions, though these remain within the acceptable range.”

May combine two claims.

The ethical mode achieves the lowest other risk overall, while the selfish mode still maintains a relatively low level of risk.

“Specifically, the ethical mode achieves the lowest other risk overall, while the selfish mode still maintains a relatively low level of risk.”

May combine two claims.

EthicAR w/o PER performed poorly, with compliance below 30% on risky steps, confirming that prioritized experience replay is essential for learning ethical behavior when violations are rare.

“EthicAR w/o PER performed poorly, with compliance below 30% on risky steps, confirming that prioritized experience replay is essential for learning ethical behavior when violations are rare.”

May combine two claims.

For EthicAR w/o PER, the scarcity of cost signals prevents the agent from learning meaningful behaviors related to cost limitation, and λ remains near zero throughout training.

“For EthicAR w/o PER, the scarcity of cost signals prevents the agent from learning meaningful behaviors related to cost limitation. As a result, λ remains near zero throughout training, and the agent fails to develop sensitivity to cost constraints.”

May combine two claims.

Both the ethical and selfish modes of EthicAR show markedly fewer critical cases than the standard mode, with the ethical mode performing best (only three isolated instances in the red squares).

“In contrast, both the ethical and selfish modes of EthicAR show markedly fewer critical cases, with the ethical mode performing best: only three isolated instances fall into the red squares.”

• AssumptionRests on an assumption: TTC is a critical indicator for assessing safety.
• AssumptionRests on an assumption: A predictive and cautious agent should be able to reduce risk in advance, leading to fewer instances where both TTC is low and other risk is high.

May combine two claims.

Across two interactive benchmarks and five random seeds, the policy decreases conflict frequency by 25-45% compared to matched task successes while maintaining comfort metrics within 5%.

“Across two interactive benchmarks and five random seeds, our policy decreases conflict frequency by 25 ∼ 45% compared to matched task successes while maintaining comfort metrics within 5%.”

May combine two claims.

Although EthicAR w/o PER and SACLAG can reduce risks compared to the standard mode, this reduction does not consistently improve with decreasing cost limits.

“Considering the other agents, although EthicAR w/o PER and SACLAG can reduce risks compared to the standard mode, this reduction does not consistently improve with decreasing cost limits.”

May combine two claims.

The LSTMSAC agent remained at floor levels (<15%), highlighting the gap between structured Safe RL and naive behavior.

“Finally, the LSTMSAC agent remained at floor levels ( &lt; 15% ), highlighting the gap between structured Safe RL and naive behavior.”

May combine two claims.

The SACLAG agent maintains a relatively low cost but still fails to meet the cost limit consistently.

“The SACLAG agent performs better in this regard, maintaining a relatively low cost, yet still fails to meet the cost limit consistently.”

May combine two claims.

SACLAG achieved moderate compliance (68~78% on all steps; 60~70% on risky steps), suggesting partial success in constraint learning but limited robustness in rare-event cases.

“SACLAG achieved moderate compliance ( 68 ∼ 78% on all steps; 60 ∼ 70% on risky steps), suggesting partial success in constraint learning but limited robustness in rare-event cases.”

May combine two claims.

EthicAR maintains a respectful distance from the cyclist throughout and refrains from overtaking, keeping its maximum other risk at 0.1 until the conclusion of the cyclist scenario.

“By contrast, EthicAR maintains a respectful distance from the cyclist throughout and refrains from overtaking, keeping its maximum other risk at a low 0.1 until the conclusion of the scenario, showing more socially acceptable, ethically aligned behavior suitable for everyday driving.”

• AssumptionRests on an assumption: TTC is a critical indicator for assessing safety.

May combine two claims.

For the LSTMSAC agent, while both the ethical and selfish modes show reduced risks to others, the ethical mode occasionally results in slightly higher other risk than the selfish mode.

“A similar pattern is observed with the LSTMSAC agent: while both the ethical and selfish modes show reduced risks to others, the ethical mode occasionally results in slightly higher other risk than the selfish mode.”

May combine two claims.

The risk-sensitive replay scheme improves learning efficiency in rare but safety-critical interactions, validated through ablation.

“We introduce a risk-sensitive replay scheme that improves learning efficiency in rare but safety-critical interactions, validated through ablation.”

May combine two claims.

EthicAR in selfish mode shows slightly higher accelerations than other agents but remains within the comfort zone.

“According to Figure 5, EthicAR in selfish mode shows slightly higher accelerations but still remains within the comfort zone.”

May combine two claims.

Even though the ego vehicle legally had right-of-way, EthicAR proactively yields to the red vehicle, keeping overall risk at a low level.

“Even though the ego vehicle legally had right-of-way, EthicAR proactively yields to the red vehicle, keeping overall risk at a low level.”

• AssumptionRests on an assumption: In the right-turn merging conflict scenario, the ego vehicle has right-of-way on the main road, and the right-turning red vehicle must yield.

May combine two claims.

The SACLAG agent shows improved restraint in the unprotected left-turn scenario: though ego speed remains relatively high at T=4s, TTC and risk increase to about 0.3, still yielding more courteously.

“The SACLAG agent shows improved restraint: though ego speed remains relatively high at T = 4s , TTC and risk increase to about 0.3, still yielding more courteously.”

• AssumptionRests on an assumption: TTC is a critical indicator for assessing safety.

May combine two claims.

The EthicAR w/o PER agent performs slightly better, its minimal deceleration increases the TTC with the red vehicle, but it still adopts a dangerously aggressive approach.

“The EthicAR w/o PER agent performs slightly better, its minimal deceleration increases the TTC with the red vehicle, but it still adopts a dangerously aggressive approach.”

• AssumptionRests on an assumption: In the unprotected left-turn conflict scenario, the ego vehicle nominally has the right of way, but an early-arriving or aggressively driven red vehicle may still pose a hazard.

May combine two claims.

In the unprotected left-turn scenario, the standard mode starts from inf TTC and risk 0, and allows the other vehicle to turn.

“For instance, for the standard mode, starting from inf TTC, and risk of 0, and then allows the other vehicle to turn.”

• AssumptionRests on an assumption: TTC is a critical indicator for assessing safety.

May combine two claims.

The ethical mode of EthicAR outperforms the other variants, significantly reducing overall risk for all road users compared to versions without cost constraints.

“Our results show that the ethical mode of EthicAR outperforms the other variants, significantly reducing overall risk for all road users compared to versions without cost constraints.”

• AssumptionRests on an assumption: TTC is a critical indicator for assessing safety.

May combine two claims.

By T=6s, both standard mode and EthicAR w/o PER collide with the red vehicle; SACLAG narrowly avoids impact but still operates in a high-risk regime.

“By T = 6s , both standard mode and EthicAR w/o PER collide with the red vehicle; SACLAG narrowly avoids impact but still operates in a high-risk regime.”

• AssumptionRests on an assumption: TTC is a critical indicator for assessing safety.
• AssumptionRests on an assumption: In the right-turn merging conflict scenario, the ego vehicle has right-of-way on the main road, and the right-turning red vehicle must yield.

May combine two claims.

In the right-turn merging conflict scenario, EthicAR behaves more cautiously: at T=1s it drifts slightly left to create space for the cyclist and settles into a comfortable speed; as the red vehicle initiates its turn, TTC remains relatively constant and risks remain low.

“In contrast, the EthicAR agent behaves more cautiously. At T = 1s , it drifts slightly left to create space for the cyclist and settles into a comfortable speed. As the red vehicle initiates its turn, the TTC remains relatively constant, and the risks remain low.”

• AssumptionRests on an assumption: TTC is a critical indicator for assessing safety.
• AssumptionRests on an assumption: In the right-turn merging conflict scenario, the ego vehicle has right-of-way on the main road, and the right-turning red vehicle must yield.

May combine two claims.

In the unprotected left-turn scenario, the standard mode's approach remains aggressive: by T=4s, TTC falls to 0.4s and risk spikes to 0.7, even though no collision occurs.

“Yet its approach remains aggressive: by T = 4s , TTC falls to just 0.4s and risk spikes to 0.7, even though no collision occurs.”

• AssumptionRests on an assumption: TTC is a critical indicator for assessing safety.

May combine two claims.

Both SACLAG and EthicAR show the strongest safety performance: they proactively reduce speed as they near the intersection, thereby minimizing risk to all road users.

“In contrast, both SACLAG and the EthicAR agent show the strongest safety performance: they proactively reduce the speed as it nears the intersection, thereby minimizing risk to all road users.”

• AssumptionRests on an assumption: In the unprotected left-turn conflict scenario, the ego vehicle nominally has the right of way, but an early-arriving or aggressively driven red vehicle may still pose a hazard.

May combine two claims.

By accounting for the safety of other road users, we not only lower their risk but also reduce the ego risk; cooperative traffic behavior benefits others and enhances the safety of the ego vehicle.

“This finding highlights a compelling insight: by accounting for the safety of other road users, we not only lower their risk but also reduce the ego risk. In other words, cooperative traffic behavior not only benefits others but also enhances the safety of the ego vehicle.”

May combine two claims.

EthicAR's behavior in the cyclist scenario is more socially acceptable and ethically aligned, suitable for everyday driving.

“By contrast, EthicAR maintains a respectful distance from the cyclist throughout and refrains from overtaking, keeping its maximum other risk at a low 0.1 until the conclusion of the scenario, showing more socially acceptable, ethically aligned behavior suitable for everyday driving.”

May combine two claims.

EthicAR maintains cost constraints under normal operation and reliably avoids rare but critical violations.

“This indicates that EthicAR not only maintains cost constraints under normal operation but also reliably avoids rare but critical violations.”

May combine two claims.

The consistency across modes highlights both the generality of these findings and the importance of combining LSTM-based architectures with dynamic PER for safety-critical, ethics-aware learning.

“The consistency across modes highlights both the generality of these findings and the importance of combining LSTM-based architectures with dynamic PER for safety-critical, ethics-aware learning.”

The quote only partly states this claim.

The ability to learn truly ethical behavior from the cost function is limited for EthicAR w/o PER and SACLAG.

“This suggests that the ability to learn truly ethical behavior from the cost function is limited.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Jerk comfort thresholds are typically below 1 m/s3, with values between 1 and 3 m/s3 often associated with reduced comfort or instability.

“In terms of jerk, comfort thresholds are typically below 1m / s 3 , with values between 1 and 3m / s 3 often associated with reduced comfort or instability [32].”

• cites [32] sourceNot checked — no full text of the source

Classified as probably not a substantial claim. May combine two claims.

The method explicitly aims to reduce risks to all road users, unlike traditional RL approaches that focus solely on performance or safety.

“Unlike traditional RL approaches, which focus solely on performance or safety, our method explicitly aims to reduce risks to all road users .”

Classified as probably not a substantial claim.

The Equality principle term discourages policies that significantly reduce total risk by disproportionately increasing the risk for specific individuals.

“This term discourages policies that significantly reduce total risk by disproportionately increasing the risk for specific individuals.”

May combine two claims.

The counterintuitive outcome of LSTMSAC's ethical mode occasionally having higher other risk indicates ineffective training and aligns with the training performance shown in Figure 2.

“This counterintuitive outcome further indicates ineffective training and aligns with the training performance shown in Figure 2.”

Classified as probably not a substantial claim.

The Bayes principle component ensures that the agent strives to reduce the overall risk in the environment, regardless of which participant bears it.

“This component ensures that the agent strives to reduce the overall risk in the environment, regardless of which participant bears it.”

Classified as probably not a substantial claim. May combine two claims.

The reward function encourages safe, efficient, and goal-oriented behavior and balances competing objectives such as maintaining speed, staying on track, and avoiding abrupt maneuvers or collisions.

“Overall, this reward function encourages safe, efficient, and goal-oriented behavior. It balances competing objectives such as maintaining speed, staying on track, and avoiding abrupt maneuvers or collisions.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The ethical cost function depends on chosen weights and harm models, which may reflect implicit normative assumptions.

“the ethical cost function depends on chosen weights and harm models, which may reflect implicit normative assumptions”

Probably not a substantial claim: may not pass S3 (load-bearing).

The protection of VRUs could be enhanced through straightforward adjustments to weight assignments.

“This outcome underscores the effectiveness of the proposed framework and further suggests that the protection of VRUs could be enhanced through straightforward adjustments to weight assignments.”

Classified as probably not a substantial claim.

The framework's goal is to balance efficiency, safety, and ethical behavior.

“Its goal is to balance efficiency, safety, and ethical behavior.”

May combine two claims.

The agent learns to reduce unnecessary proximity to vulnerable individuals or to distribute unavoidable risk more equitably.

“For example, the agent learns to reduce unnecessary proximity to vulnerable individuals or to distribute unavoidable risk more equitably.”

Probably not a substantial claim: may not pass S2 (checkable).

Focusing on extreme dilemmas risks obscuring the everyday ethical trade-offs AVs face, such as yielding to pedestrians at crosswalks, balancing risk across multiple agents, and ensuring equitable safety for VRUs.

“However, focusing on such extreme dilemmas risks obscuring the everyday ethical trade-offs AVs face, such as yielding to pedestrians at crosswalks, balancing risk across multiple agents, and ensuring equitable safety for VRUs.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The probabilistic risk model's independence assumption may underestimate compounding risk.

“our probabilistic risk model assumes independence across timesteps, which may underestimate compounding risk.”

• AssumptionThe authors state this as an assumption: Collisions at different timesteps are independent.

May combine two claims.

The method enables efficient and accurate risk estimation, avoiding unnecessary computation when objects are clearly not in proximity while still capturing nuanced spatial uncertainty when they are.

“This method enables efficient and accurate risk estimation. It avoids unnecessary computation when objects are clearly not in proximity, while still capturing nuanced spatial uncertainty when they are.”

Classified as probably not a substantial claim.

By incorporating the ethical cost into the constraint, planned trajectories reflect not only collision avoidance but also a broader sense of moral responsibility.

“By incorporating this ethical cost, described in detail in Section 3.5, directly into the constraint, our approach ensures that planned trajectories reflect not only a commitment to collision avoidance but also a broader sense of moral responsibility.”

Probably not a substantial claim: may not pass S4 (based).

The planner generates collision-free trajectories using the RL outputs.

“Using these outputs, the planner generates collision-free trajectories”

May combine two claims.

The evaluation uses 75 unseen, real-world scenarios drawn from the Waymo dataset, held out during training to ensure an unbiased comparison.

“To further assess the performance of the proposed agent and baseline methods, we evaluate all models on 75 unseen, real-world scenarios drawn from the Waymo dataset. These scenarios were held out during training to ensure an unbiased comparison.”

May combine two claims.

Overall compliance can appear inflated due to the predominance of trivially safe steps, whereas risky-step compliance better reflects an agent's ability to enforce constraints in critical situations.

“A key observation is that overall compliance can appear inflated due to the predominance of trivially safe steps, whereas risky-step compliance better reflects an agent's ability to enforce constraints in critical situations.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

This ethical integration sets the method apart from conventional AV planning techniques, which often treat safety and ethics as secondary considerations.

“This ethical integration sets our method apart from conventional A V planning techniques, which often treat safety and ethics as secondary considerations.”

May combine two claims.

The agent does not receive predicted trajectories as input, but by leveraging its LSTM architecture with the current states of nearby vehicles provided as input, it implicitly learns to model these predictions through cost-based training.

“While the agent does not receive these predicted trajectories as input, by leveraging its LSTM architecture with the current states of nearby vehicles provided as input, the agent implicitly learns to model these predictions through cost-based training.”

• AssumptionRests on an assumption: A predictive and cautious agent should be able to reduce risk in advance, leading to fewer instances where both TTC is low and other risk is high.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The observation and action space are designed to ensure compatibility with dynamic road geometries.

“The observation and action space, detailed in Section 4.2, are designed to ensure compatibility with dynamic road geometries.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The execution level translates high-level targets into smooth, feasible trajectories, ensuring both accuracy and comfort.

“At the execution level, polynomial path planning coupled with Proportional-Integral-Derivative (PID) and Stanley controllers translates these targets into smooth, feasible trajectories, ensuring both accuracy and comfort.”

Probably not a substantial claim: may not pass S4 (based).

The transitional period will involve extensive interaction between AVs and human-driven vehicles.

“the transitional period will involve extensive interaction between A Vs and human-driven vehicles.”

May combine two claims.

The suitable operating range for cost threshold η is [0.6, 1]; thresholds below 0.6 are unrealistically strict, and above 1 are too permissive.

“From Figure 3, we identify η ∈ [0 . 6 , 1] as the suitable operating range. Thresholds below 0.6 are unrealistically strict, such that almost no agent can comply, while thresholds above 1 are too permissive to meaningfully enforce ethical behavior.”

Classified as probably not a substantial claim.

The policy evolves with a balanced emphasis on both task performance and ethical responsibility.

“As a result, the policy evolves with a balanced emphasis on both task performance and ethical responsibility.”

May combine two claims.

The framework is trained and evaluated in closed-loop simulation environments derived from large-scale, real-world traffic datasets with heterogeneous road users, enabling reproducible and realistic evaluation across complex scenarios.

“We train and evaluate EthicAR in closed-loop simulation environments derived from large-scale, real-world traffic datasets with heterogeneous road users, enabling reproducible and realistic evaluation across complex scenarios.”

May combine two claims.

EthicAR successfully enforces ethical constraints even under rare-event conditions, while SACLAG only partially succeeds and alternative approaches fail.

“Taken together, these results show that EthicAR successfully enforces ethical constraints even under rare-event conditions, while SACLAG only partially succeeds and alternative approaches fail.”

• AssumptionThe authors state this as an assumption: A reliable perception module exists that provides accurate, real-time information about surrounding traffic participants.

May combine two claims.

The Waymo Open Dataset contains recordings from six cities across the United States and includes a wide variety of challenging scenarios such as unprotected turns, merges, lane changes, and complex intersection interactions.

“For the real-world data, we utilize the Waymo Open Dataset [27], which contains recordings from six cities across the United States. It includes a wide variety of challenging scenarios, such as unprotected turns, merges, lane changes, and complex intersection interactions, making it particularly well-suited for our task.”

• cites [27] sourceNot checked — no full text of the source

May combine two claims.

The two-level control architecture separates high-level decision making (Safe RL) from low-level execution (classical controllers), ensuring both principled reasoning and stable motion.

“Our framework separates high-level decision making (Safe RL) from low-level execution (classical controllers), ensuring both principled reasoning and stable motion.”

Classified as probably not a substantial claim.

The framework's ethical integration is a key contribution that distinguishes it from conventional AV planning.

“A key contribution of our framework is the integration of ethical risk minimization into the RL objective.”

May combine two claims.

This enables EthicAR to exhibit predictive behavior, resulting in higher TTC values and ultimately lower risk during interactions with other traffic participants.

“This enables EthicAR to exhibit predictive behavior, resulting in higher TTC values and ultimately lower risk during interactions with other traffic participants.”

• AssumptionRests on an assumption: TTC is a critical indicator for assessing safety.
• AssumptionRests on an assumption: A predictive and cautious agent should be able to reduce risk in advance, leading to fewer instances where both TTC is low and other risk is high.

May combine two claims.

Forcing an overtake on a narrow, two-lane road with oncoming congestion violates both ethical norms and common etiquette and risks a collision with the cyclist or opposing traffic due to cyclist unpredictability.

“In reality, forcing an overtake on a narrow, two-lane road with oncoming congestion violates both ethical norms and common etiquette [33], and risks a collision either with the cyclist or opposing traffic due to the unpredictability of the cyclist.”

• cites [33] sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

The ethical mode cost function accounts for the risks imposed on all traffic participants, encouraging decisions that are socially responsible.

“In the ethical mode, the cost function accounts for the risks imposed on all traffic participants, encouraging decisions that are socially responsible.”

May combine two claims.

Compared to Monte Carlo sampling, the approach offers substantial computational savings without sacrificing predictive reliability.

“Compared to Monte Carlo sampling, our approach offers substantial computational savings without sacrificing predictive reliability.”

Classified as probably not a substantial claim. May combine two claims.

The Maximin principle distinguishes between frequent low-severity events and rare but potentially catastrophic outcomes, ensuring that decisions are sensitive to the most vulnerable participants.

“This principle distinguishes between frequent low-severity events and rare but potentially catastrophic outcomes, ensuring that decisions are sensitive to the most vulnerable participants.”

Probably not a substantial claim: may not pass S4 (based).

The cost functions promote fair and responsible risk distribution across all participants.

“we design cost functions that promote fair and responsible risk distribution across all participants.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Frenet coordinate system simplifies trajectory planning on curved roads by decoupling progress and deviation.

“The RL agent operates within a Frenet coordinate system, which describes motion relative to a reference path using two parameters: l , the longitudinal distance along the path, and d , the lateral offset from the path. This representation simplifies trajectory planning on curved roads by decoupling progress and deviation, providing an intuitive view of the vehicle's relationship to the road.”

May combine two claims.

Both the LSTM architecture and dynamic PER play a critical role in enabling effective learning under explicit cost constraints.

“These findings highlight the critical role of both the LSTM architecture and dynamic PER in enabling effective learning under explicit cost constraints.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The evaluation relies on dataset-derived closed-loop simulations, where surrounding agents are only partially reactive; fully interactive multi-agent dynamics require further validation.

“our evaluation relies on dataset-derived closed-loop simulations, where surrounding agents are only partially reactive; fully interactive multi-agent dynamics require further validation.”

Probably not a substantial claim: may not pass S4 (based).

The composite ethical risk cost integrates collision probability and harm severity, enabling Safe RL to balance ego safety with fairness to other road users.

“We design a composite ethical risk cost that integrates collision probability and harm severity, enabling Safe RL to balance ego safety with fairness to other road users.”

May combine two claims.

The case studies confirm that EthicAR consistently makes responsible decisions, not only toward other vehicles but also toward VRUs such as cyclists.

“These case studies confirm that EthicAR consistently makes responsible decisions, not only toward other vehicles but also toward VRUs such as cyclists.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

If stronger prioritization of VRUs is required, this can be readily achieved through simple weight adjustments in the cost function.

“If stronger prioritization of VRUs is required, this can be readily achieved through simple weight adjustments in the cost function.”

May combine two claims.

EthicAR not only reacts ethically in high-risk situations but also incorporates ethical considerations into routine driving behavior, thereby proactively avoiding dangerous scenarios.

“EthicAR not only reacts ethically in high-risk situations but also incorporates ethical considerations into routine driving behavior, thereby proactively avoiding dangerous scenarios.”

May combine two claims.

The standard mode's aggressive approach in the unprotected left-turn scenario is risky and unsuitable for AV deployment.

“Such behavior is risky and unsuitable for A V deployment.”

Classified as probably not a substantial claim.

The trained agents are tested on 75 previously unseen real-world scenarios.

“For performance evaluation, we test the trained agents on 75 previously unseen real-world scenarios.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The agents are trained using a real-world dataset in simulation, providing realistic scenarios that improve generalization and reliability.

“The agents are trained using a real-world dataset in simulation, providing realistic scenarios that improve generalization and reliability.”