PedSLight: pedestrian and safety-aware reinforcement learning for mixed pedestrian-vehicle signal control

1 Introduction

Signalized intersections are critical nodes where left-turning vehicles and crossing pedestrians interact intensively (Fu et al., 2026b). Although conventional signal control primarily targets vehicular efficiency, pedestrian safety remains a major concern under permissive left-turn operations (Qi and Guoguo, 2017; Chen et al., 2015). Field conflict studies further show that left-turning movements account for a substantial share of pedestrian–vehicle interactions at signalized intersections and that conflict rates vary with intersection geometry (Lord, 1996). Conflicts between left-turn vehicles and pedestrians constitute a substantial portion of severe intersection crashes, motivating safety-aware control strategies tailored to turning movements.

Time-to-collision (TTC) is widely adopted for proactive conflict assessment (Saffarzadeh et al., 2013), dating back to early near-miss formulations that quantify remaining time to a potential collision under constant-speed trajectories (Hayward, 1972). Subsequent work has compared TTC with related indicators such as time headway (Vogel, 2003) and formalized surrogate safety measures extractable from microscopic simulation, including TTC and post-encroachment time (Gettman and Head, 2003). Yet TTC is rarely embedded directly into adaptive signal optimization. Existing formulations typically optimize aggregate delay or incorporate safety through structural phase adjustments, limiting the ability to represent interaction risk explicitly within the control objective.

Reinforcement learning (RL) has demonstrated strong performance in adaptive signal control, often outperforming fixed-time and actuated strategies in delay reduction (Poudel et al., 2025; Aslani et al., 2017). Model-based and optimization-oriented approaches have also introduced pedestrian-protection mechanisms through logical constraints and phase design (Zhang et al., 2021). Recent extensions incorporate pedestrian considerations through phase restructuring or rule-based constraints (Xu et al., 2024; Zhang et al., 2025; Ren et al., 2025; Du et al., 2023; Barhoumi et al., 2025). However, safety is often incorporated through phase design or constraint-based mechanisms rather than modeled as a risk term directly embedded in the learning objective. In addition, RL formulations that update actions at short simulation intervals may result in frequent phase adjustments, potentially affecting operational stability.

Beyond isolated intersections, multi-agent reinforcement learning (MARL) has become a dominant paradigm for coordinated adaptive traffic signal control (ATSC) over arterial and grid networks. Early multi-agent formulations such as MARLIN-ATSC demonstrated coordinated learning over large urban networks (El-Tantawy et al., 2013), and deep RL methods such as IntelliLight established learnable phase-selection policies at individual signals (Wei et al., 2018). PressLight and CoLight extended MARL to network-level coordination by aligning rewards with traffic pressure and enabling attention-based communication among neighboring intersections (Wei et al., 2019a; b), while multi-agent actor–critic methods further stabilize decentralized training under partial observability (Chu et al., 2019). Scalable pressure-based designs such as MPLight have been evaluated on city-scale networks with thousands of signals (Chen et al., 2020). CVLight further advances decentralized MARL for connected-vehicle environments through an asymmetric actor–critic design that supports scalable multi-intersection control under partial observability (Mo et al., 2022). More recently, hierarchical federated reinforcement learning has been proposed to group heterogeneous intersections and coordinate policy learning with limited raw-data exchange across agents (Fu et al., 2026a). Although these MARL frameworks substantially improve vehicular efficiency and scalability, pedestrian conflict risk is rarely incorporated explicitly into the reward design of multi-agent traffic signal control.

To address these limitations, we propose PedSLight, a cycle-level reinforcement learning framework for mixed vehicular–pedestrian signal control that integrates a TTC-based safety metric directly into the reward formulation. The problem is formulated using a safety-aware weighted optimization structure to enable controllable efficiency–safety trade-offs. Decisions are updated at cycle boundaries to ensure phase stability while adaptively allocating green time between protected and permissive left-turn movements.

The main contributions of this work can be summarized as follows:

• We propose a cycle-level reinforcement learning framework for mixed vehicular–pedestrian signal control and incorporate a TTC-based surrogate metric to represent pedestrian safety risk.
• We formulate signal control as a safety-aware weighted optimization problem, which provides explicit control over the efficiency–safety trade-off during training.
• We conduct a demand-dependent performance analysis under multiple traffic regimes and show that adaptive control exhibits more robust behavior as traffic intensity increases.

The remainder of this paper is organized as follows. Section 2 reviews related work on pedestrian safety and learning-based signal control. Section 3 establishes the cycle-level control formulation for mixed vehicular–pedestrian intersections. Section 4 presents the proposed cycle-level reinforcement learning approach and details its implementation and training procedure. Section 5 evaluates the approach against conventional baselines under varying traffic demand conditions. Finally, Section 6 concludes the study.

2 Related work

Existing studies on safety-aware traffic signal control at signalized intersections can be broadly categorized into the following three directions.

2.1 Pedestrian conflict analysis and surrogate safety modeling

Simulation and quasi-experimental studies indicate that pedestrian volume, heavy-vehicle presence, and left-turn phasing are associated with pedestrian–vehicle conflicts, and that the observed safety effects of protected or protected/permissive control may vary across traffic and geometric conditions (Qi and Guoguo, 2017; Chen et al., 2015). Surrogate safety measures are frequently adopted to characterize conflict proximity. However, in signal-control studies they are more often used for safety assessment than incorporated directly into the control objective.

2.2 Pedestrian-aware and model-based control

Pedestrian-aware control has been addressed through analytical and optimization-based approaches. Dynamic all-red phases and logical constraints have been introduced to enhance pedestrian protection (Zhang et al., 2021), and Ped-MP incorporates pedestrian access into the max-pressure framework while preserving network stability (Xu et al., 2024). At isolated intersections, joint vehicle–pedestrian timing has also been formulated as a quadratic program that minimizes weighted delays subject to crosswalk and refuge-island constraints (Yu et al., 2017), while integrated models jointly select pedestrian phase patterns and signal timings using delay, exposure, and violation measures (Ma et al., 2015). These approaches provide transparent operational constraints, but generally rely on predefined demand and behavioral models rather than an adaptive policy learned from evolving traffic states.

2.3 Learning-based control and safety considerations

Deep reinforcement learning (DRL) has been applied to jointly optimize vehicular and pedestrian efficiency, primarily through delay and waiting-time minimization (Poudel et al., 2025), as well as to mitigate conflicts via protected/prohibited turning strategies (Zhang et al., 2025; Ren et al., 2025). A parallel stream embeds predicted crash risk in multi-objective ATSC: real-time crash models have been combined with mobility objectives in multi-objective RL (Gong et al., 2020; Reyad and Sayed, 2023), and crash risk alone has been optimized in real time (Reyad et al., 2022). More recent work expands the objective set to safety, efficiency, and emissions using D3QN (Zhang et al., 2024), or compares PPO, A2C, and DQN in both isolated and corridor settings (Elharoun et al., 2025). Cycle-level crash probabilities estimated from video-derived conflicts have likewise been coupled with a DQN controller (Tahir and Haque, 2026). The crash-risk and broad multi-objective studies in this latter stream principally concern vehicle–vehicle crash risk or network-wide conflict counts; they do not isolate the geometry and directional pairing of permissive-left-turn vehicle–pedestrian encounters addressed here.

Safety-aware RL can instead incorporate safety standards or impose explicit constraints. SafeLight integrates road safety standards through a safety-enhanced residual RL design with multi-objective loss and reward shaping (Du et al., 2023). Formally constrained controllers check proposed actions against specified safety conditions (Barhoumi et al., 2025), while constrained RL with a learned cost estimator and Lagrangian updates has been used to limit unsafe phase switching under competing transit-priority requests (Zhou et al., 2025). These approaches emphasize rule enforcement or constraint satisfaction rather than optimizing a cycle-aggregated pedestrian-conflict surrogate. Separately, video-based LSTM models can predict pedestrian–vehicle conflicts in advance (Zhang et al., 2020), but prediction alone does not determine signal actions.

Table 1 positions PedSLight against the studies highlighted above. The novelty is not the general use of a weighted multi-objective reward; rather, it is the combination of (i) a direction- and geometry-aware TTC surrogate specifically for permissive-left-turn vehicle–pedestrian interactions, (ii) continuous allocation between protected and permissive service while retaining a fixed phase sequence, (iii) cycle-boundary actions that avoid rapid phase switching, and (iv) one formulation applied to both an isolated intersection and a decentralized corridor.

TABLE 1. Comparison with related safety- and pedestrian-aware signal-control studies.

3 Problem formulation

We consider a network of N signalized intersections with mixed vehicular and pedestrian traffic. Each intersection i ∈ I = { 1 , … , N } is modeled as a four-leg junction with four approaches (North, East, South, West), each including one through lane, one left-turn lane, and an associated pedestrian crossing. All intersections operate under the same fixed signal structure consisting of four green phases and two fixed yellow transitions, with cycle-based decision updates. When N = 1, the formulation reduces to an isolated intersection; when N > 1, the network is controlled by a decentralized multi-agent system in which each intersection is assigned an independent controller that observes local traffic states and selects local timing actions.

At the beginning of each signal cycle, controller i determines the green time allocation between the protected left-turn phase and the permissive left-turn phase for the North–South and East–West directional groups at intersection i. The selected allocation remains fixed throughout the entire cycle at that intersection. Intersection coupling in multi-intersection settings arises through shared corridor traffic dynamics (e.g., queue spillback and progression effects), but each agent conditions its decision only on locally observable states and receives a local cycle-level reward.

The control objective is to balance traffic efficiency and pedestrian safety under dynamic traffic demand across the network.

The overall cycle-level safety-aware reinforcement learning framework is illustrated in Figure 1. At each cycle boundary, every controller observes a direction-aware local state tensor, selects a continuous green-time allocation action, and receives a cycle-level reward that integrates both efficiency and TTC-based safety cost.

3.1 State construction

At decision step t, the local environment state at intersection i is represented using a Discrete Traffic State Encoding (DTSE), formulated as a direction-aware tensor in Equation 1:as illustrated in Figure 1.

s t i ∈ R 12 × 10 × 2 , (1)

The first dimension corresponds to 12 movement groups, formed by four approaches (North, East, South, and West), each containing three streams: pedestrian, through, and left-turn movements.

The second dimension discretizes the longitudinal space into 10 non-uniform distance bins. Finer resolution is assigned near the intersection to better capture pedestrian-vehicle interaction dynamics.

The third dimension contains two feature channels: a binary occupancy indicator and the average speed of agents within each cell. Unoccupied cells contain zero values.

Collecting all intersections yields the joint system state in Equation 2:In the decentralized setting considered here, controller i uses only s t i as its input at cycle t.

s t = s t 1 , s t 2 , … , s t N . (2)

3.2 Action space

The signal control action is defined at the cycle level for each intersection. At the beginning of each signal cycle, controller i selects a continuous action in Equation 3:where α NS i , α EW i ∈ [ 0,1 ] denote the green time allocation ratios for the North–South and East–West directional groups at intersection i, respectively.

a t i = α NS i , α EW i , (3)

Each parameter α ∈ [ 0,1 ] determines the proportion of green time assigned to the protected left-turn phase within the corresponding directional group, as illustrated in Figure 2. The remaining green time is allocated to the permissive left-turn phase.

The selected action remains fixed throughout the entire cycle and is updated only at the next cycle boundary. This cycle-based decision structure ensures phase stability while allowing adaptive adjustment of left-turn protection intensity under dynamic traffic conditions at each intersection.

The joint action profile at cycle t is denoted by Equation 4.

a t = a t 1 , a t 2 , … , a t N . (4)

3.3 Efficiency and safety definitions

At decision epoch t, let V t i and P t i denote the numbers of vehicles and pedestrians currently present within the control zone of intersection i, respectively. Traffic efficiency at intersection i is measured by the average waiting time of all active agents in that local zone at the decision epoch as defined in Equation 5:where V t i and P t i denote the sets of active vehicles and pedestrians at intersection i, and W v ( t ) and W p ( t ) denote their accumulated waiting times. The instantaneous local efficiency reward is defined in Equation 6:

W ̄ t i = 1 V t i + P t i ∑ v ∈ V t i W v t + ∑ p ∈ P t i W p t , (5)

E t i = − W ̄ t i . (6)

Pedestrian safety at intersection i is quantified by a time-to-collision (TTC) based surrogate cost C t i that captures potential conflicts between left-turn vehicles and conflicting pedestrian movement streams within the local control zone. The detailed computation of C t i is described in Section 4.

For network-level analysis, local efficiency and safety terms can be aggregated across intersections as in Equation 7, e.g.,although each controller is trained using its own local reward unless otherwise specified.

E t = 1 N ∑ i = 1 N E t i , C t = 1 N ∑ i = 1 N C t i , (7)

3.4 Safety-aware optimization objective

We consider a safety-aware optimization objective that balances traffic efficiency and pedestrian safety at each intersection. The combined local reward is defined in Equation 8:where λ ≥ 0 controls the relative emphasis on safety. In the decentralized multi-agent setting, each intersection i is associated with a policy π θ i (or a shared policy π θ with parameter sharing) that maps local states to local actions, i.e., a t i ∼ π θ i ( ⋅ ∣ s t i ). The learning objective for agent i is to maximize the expected discounted return in Equation 9:where γ ∈ ( 0,1 ) is the discount factor. When policies are shared across intersections, the objective is to maximize the average expected return over agents in Equation 10:This formulation supports both isolated single-intersection control ( N = 1 ) and decentralized multi-intersection deployment ( N > 1 ) within the same cycle-level safety-aware framework.

R t i = E t i − λ C t i , (8)

J i π θ i = E π θ i ∑ t = 0 ∞ γ t R t i , (9)

J π θ = 1 N ∑ i = 1 N J i π θ . (10)

The weighted objective is a safety-aware formulation, but it should not be interpreted as a hard safety guarantee. Safe DRL encompasses distinct mechanisms: constrained Markov decision process methods can enforce expected cost budgets through Lagrangian updates (Zhou et al., 2025), formal controllers can check proposed actions against specified safety conditions (Barhoumi et al., 2025), and safety-enhanced residual RL can incorporate road safety standards into the learning process (Du et al., 2023). By contrast, the present reward penalizes expected TTC-surrogate exposure, allowing a tunable empirical trade-off through λ but not guaranteeing that every action is conflict-free. This choice is deliberate because near-conflict exposure is count-based and demand-dependent, and the experiments are intended to assess policy behavior in simulation. For field deployment, the learned policy should be placed beneath a runtime-assurance layer that checks minimum greens, clearance intervals, mutually exclusive movements, and a conservative conflict rule before execution, replacing an unsafe proposal with a verified fallback timing plan. The current minimum-green and fixed-sequence logic enforces operational feasibility, but it is not a complete runtime-assurance monitor; integrating and formally validating such a monitor is therefore left for deployment-oriented work.

4 Methodology

To operationalize the optimization problem defined in Section 3, we implement a cycle-level reinforcement learning framework within a microscopic traffic simulator. This section describes how local control actions are translated into executable signal plans at each intersection and how the simulator is interfaced at each decision epoch.

4.1 Cycle-level signal implementation

The action formulation is defined in Section 3. This subsection describes its execution within the simulator.

At each decision epoch, controller i updates the active traffic signal program at intersection i by reconstructing the phase durations of the corresponding SUMO signal logic. Each signal follows a fixed sequence consisting of four green phases (protected and permissive movements for the North–South and East–West directional groups) and two fixed yellow transitions.

Based on the selected decision a t i, the green durations of the protected and permissive phases are computed and assigned to the corresponding phases in the sequence. The two yellow intervals remain unchanged.

A minimum green duration constraint is imposed on both protected and permissive phases to ensure operational feasibility and to avoid impractically short signal intervals.

The updated signal logic is written back to the simulator via TraCI and remains fixed for the entire cycle at intersection i. The simulator advances sequentially until all active cycles complete. Control decisions are updated only at cycle boundaries. When N > 1, these updates are executed independently and in parallel across intersections within the same simulation step.

4.2 Reward computation

At each signal cycle t, controller i receives a local efficiency signal and a local safety signal derived from traffic states within its control zone.

To capture intra-cycle dynamics while maintaining numerical stability, the efficiency signal is evaluated at two representative instants within the cycle (mid-point and cycle end) and averaged. In contrast, the safety signal is defined as a cycle-wise TTC-based conflict cost accumulated over all simulation steps within the cycle at intersection i.

4.2.1 Snapshot-based efficiency computation

Let E t i , m i d and E t i , e n d denote the instantaneous local efficiency measurements obtained at the mid-point and end of the signal cycle at intersection i, respectively.

Each measurement is computed using the current accumulated waiting times of all active agents (vehicles and pedestrians) within the local control zone of intersection i.

Since the simulator continuously updates waiting time when an agent’s speed remains below a small threshold, the snapshot directly reflects real-time congestion conditions.

The cycle-level local efficiency signal is defined in Equation 11:

E t i = 1 2 E t i , m i d + E t i , e n d . (11)

This two-point evaluation reduces sensitivity to short transient fluctuations caused by phase switching while preserving responsiveness to signal timing decisions. Lower delay corresponds to higher efficiency as defined in Section 3. Therefore, E t i is negatively correlated with congestion level at intersection i.

4.2.2 TTC-based safety computation

Pedestrian safety at intersection i is evaluated using a threshold-based time-to-collision (TTC) surrogate that detects near-simultaneous arrivals of left-turn vehicles and geometrically conflicting pedestrian streams within the local control zone.

Only vehicles in left-turn lanes are considered as potential conflict sources. For each left-turn vehicle v and pedestrian p in the conflicting stream at intersection i, the estimated arrival times to the conflict boundary are in Equation 12 where d v and d p denote the projected distances to the conflict region, v v and v p are the instantaneous speeds, and ϵ v , ϵ p > 0 ensure numerical stability.

t v = d v max v v , ϵ v , t p = d p max v p , ϵ p , (12)

To reflect intersection geometry, pedestrian candidates are filtered by movement direction consistent with the turning conflict, and the spatial location on the turn-in side of the intersection.

A potential conflict event is recorded when the arrival times satisfy condition in Equation 13:where τ is a predefined TTC threshold.

The local safety cost of cycle t at intersection i is defined as the total number of such near-conflict events accumulated over the entire signal cycle in Equation 14:where k indexes simulation steps within the cycle and U i denotes the set of vehicle–pedestrian pairs evaluated within the control zone of intersection i. This definition yields a discrete near-conflict count per cycle, serving as a surrogate indicator of interaction intensity rather than a continuous risk magnitude.

C t i = ∑ k ∈ cycle t ∑ v , p ∈ U i 1 | t v k − t p k | < τ , (14)

4.2.3 Final reward construction

The cycle-level reward is assembled according to the safety-aware objective defined in Section 3, using the snapshot-based efficiency signal E t i and the cycle-wise TTC-based surrogate conflict cost C t i computed above. The resulting reward is returned to the corresponding PPO agent at the end of each signal cycle. For network-level monitoring, the local efficiency and surrogate conflict terms are aggregated across intersections as described in Section 3.

4.3 Training strategy

4.3.1 Policy optimization

We adopt Proximal Policy Optimization (PPO) (Schulman et al., 2017). The action a t i = ( α NS i , α EW i ) lies in a bounded two-dimensional continuous space, so a stochastic PPO actor can represent the allocation ratios directly. A DQN (Mnih et al., 2015) would require discretizing both ratios; with K admissible values per direction, the joint action set grows to K 2, and coarse bins reduce timing resolution while fine bins increase value-estimation and exploration burden. TD3 (Fujimoto et al., 2018) also supports continuous actions and can be more sample-efficient because it is off-policy, but its deterministic actor explores through added action noise and depends on replay-buffer data collected under older policies. In this problem, demand is randomized by episode, multi-agent corridor dynamics are non-stationary from each local agent’s perspective, and the TTC penalty can be sparse and discontinuous. We therefore favor PPO’s on-policy updates and stochastic policy, accepting lower sample efficiency in exchange for simpler bounded-action exploration and reduced sensitivity to stale replay data. Its clipped likelihood-ratio objective further limits abrupt policy changes, which is desirable because each action controls a full signal cycle. This is a problem-structure rationale rather than a claim that PPO is universally superior; a controlled PPO–TD3–DQN benchmark remains useful future work. The overall training procedure is summarized in Algorithm 1. For N = 1, we utilize Stable-Baselines3 to implement a single-agent framework; for N > 1, each intersection is assigned an independent PPO agent through a multi-agent interface (RLlib), while retaining the same local state, action, and reward structure.

4.3.2 Training demand regime

Training is conducted under a medium-flow demand scenario. To enhance robustness and avoid overfitting to a single traffic pattern, each episode applies a ± 20 % random perturbation to vehicle and pedestrian demand levels.

4.3.3 Hyperparameter configuration

Unless otherwise specified, PPO hyperparameters follow the default settings of Stable-Baselines3 and are summarized in Table 2.

TABLE 2. PPO training hyperparameters.

Algorithm 1 Training procedure for the proposed decentralized safety-aware signal controller.

5 Experiments

5.1 Setup

Experiments are conducted under two demand groups representing different traffic intensity regimes: a low-to-medium flow group and a medium-to-high flow group. Each demand group contains two fixed test profiles to capture variability within the same congestion level.

Stochastic flow perturbation is disabled during evaluation to ensure consistent traffic conditions.

For every controller-test pair, five independent simulation runs are conducted using different random seeds, each lasting 100 signal cycles. Each signal cycle is fixed at 100 s, consisting of four adjustable green phases and two 3-s yellow intervals. The single-intersection experiments use N = 1, whereas the corridor experiments use N = 3 under the same local control formulation.

Performance metrics are first averaged across the five runs for each test profile, then averaged within each demand group, and finally compared across demand regimes. Averaging across seeds mitigates stochastic variability in microscopic simulation, while evaluation across multiple demand groups improves robustness to varying traffic intensities.

5.2 Baselines

We compare the proposed safety-aware controller, PedSLight, against two conventional signal control strategies.

5.2.1 Actuated control

An actuated signal controller implemented in SUMO is adopted as a reactive baseline. The controller adjusts green durations according to real-time detector measurements under predefined phase constraints. For fair comparison, performance metrics are evaluated over the same cycle horizon as the learning-based controller, while the internal signal timings remain dynamically determined by the actuated logic.

5.2.2 Fixed-time control

A symmetric fixed-time controller is implemented by assigning identical green time allocation ratios α to both directional groups. The parameter is evaluated over α ∈ { 0.0 , 0.1 , 0.2 , 0.3 , 0.4 , 0.5 , 0.6 } to cover a range of balanced and moderately biased timing configurations. Values beyond this range are not considered to avoid excessively unbalanced green allocations.

5.3 Evaluation metrics

Two system-level performance metrics are adopted for evaluation: average travel time and TTC-based safety cost.

5.3.1 Average travel time

Let N denote the number of vehicles generated during an evaluation episode. The average travel time is defined in Equation 15:where t v entry and t v exit denote the network entry and exit times of vehicle v, respectively. Demand generation is completed within the pre-defined episode, and the resulting throughput is verified to be identical across controllers. Therefore, differences in average travel time reflect control performance rather than incomplete demand clearance.

A T T = 1 N ∑ v t v exit − t v entry . (15)

While snapshot-based efficiency is used during training to facilitate stable learning, average travel time provides a direct system-level measure of operational efficiency during evaluation.

5.3.2 TTC-based safety cost

Safety performance is evaluated using the TTC-based near-conflict cost defined in Section 4. The reported safety metric is the average cycle-level cost over the evaluation horizon in Equation 16:where T denotes the number of signal cycles in an episode.

C ̄ = 1 T ∑ t = 1 T C t , (16)

The evaluation horizon is sufficiently long to allow demand clearance. Accordingly, the safety metric represents the average cycle-level interaction intensity over the full episode.

5.4 Single-intersection results and analysis

To study the effect of safety emphasis during training, we trained PedSLight under a range of safety weight values. Based on validation performance, we select λ = 1.0 for comparison with conventional baselines.

Table 3 summarizes the averaged results under both demand regimes.

TABLE 3. Performance comparison under different demand regimes.

5.4.1 Low-to-medium demand

Under moderate traffic intensity, the actuated controller achieves the lowest travel time. Fixed-time controllers exhibit a clear efficiency–safety trade-off: increasing α reduces safety cost but substantially increases delay. PedSLight maintains competitive travel time while achieving moderate safety cost, which indicates balanced operational behavior without extreme parameter tuning.

5.4.2 Medium-to-high demand

When congestion intensifies, performance differences become more pronounced. Fixed-time strategies show strong sensitivity to parameter selection, and large α values significantly increase travel time. In contrast, PedSLight remains competitive in both efficiency and safety metrics.

Since average travel time and safety cost are measured in different units, a direct comparison does not fully reflect the underlying trade-off. To provide a unified evaluation perspective, we define a weighted scorewhere β denotes the relative emphasis on safety. The score is computed based on the aggregated results in Table 3 and analyzed over varying β to reveal the trade-off structure among controllers.

S c o r e β = − A T T − β C ̄ ,

Figure 3 illustrates how controller rankings evolve as the safety emphasis parameter β increases. The horizontal axis represents the relative weight assigned to safety cost, while the vertical axis reports the weighted score S c o r e β. Since the score linearly combines efficiency and safety, the slope of each curve reflects the magnitude of safety cost: steeper downward trends indicate larger conflict cost.

Across both demand regimes, PedSLight and actuated controllers remain near the upper envelope across most β values, consistently ranking among the top-performing configurations. By contrast, fixed-time performance shifts noticeably as α changes, which implies a strong dependence on the preset signal split.

Under low-to-medium demand, when β is small and efficiency dominates, the actuated controller achieves the highest score due to its lowest travel time. In this regime, PedSLight is slightly inferior, suggesting that under lighter traffic conditions the advantage of adaptive learning is less pronounced. However, as β increases and safety becomes more emphasized, PedSLight improves its relative ranking and achieves competitive or superior overall scores, showing a lower TTC-based surrogate conflict cost without substantial efficiency degradation.

Under medium-to-high demand, PedSLight already ranks near the top when β is small, reflecting strong efficiency under congested conditions without substantial sacrifice in safety. As β increases, its relative position remains stable, with no noticeable performance deterioration. This pattern suggests that the learned controller achieves a balanced trade-off between delay and safety under higher traffic demand.

In summary, PedSLight maintains competitive performance across safety preferences, with clearer advantages under medium-to-high demand. This aligns with intuition: safety-aware decision-making becomes more critical as traffic interactions intensify.

5.5 Corridor evaluation

To evaluate the multi-intersection case ( N = 3 ), we apply the decentralized formulation in Section 3 to a three-intersection arterial corridor implemented in SUMO. Each intersection is controlled by an independent PPO agent through RLlib’s multi-agent interface, using the same local DTSE state representation, cycle-level action space, and TTC-based reward structure. Corridor traffic dynamics introduce coupling between neighboring intersections through queue spillback and offset-dependent progression, even though each agent remains decentralized.

Corridor experiments were conducted under relatively dense demand to expose inter-intersection interaction effects. Vehicle flows were configured at approximately 350–450 vehicles per hour and pedestrian flows at approximately 180–220 pedestrians per hour per direction, depending on the route setting. Because multi-agent training exhibited higher reward variance than the single-intersection SB3 implementation, corridor policy learning used more conservative PPO update settings, including reduced learning rates, KL regularization, and entropy regularization.

5.5.1 Fixed-allocation and offset sensitivity

Before policy learning, fixed-α sweeps and offset sensitivity analyses were conducted under deterministic corridor traffic. The fixed-α sweep showed that corridor performance was highly sensitive to local signal allocation: aggressive protected-turn settings could rapidly degrade corridor-wide rewards through congestion propagation, even when safety costs changed only moderately. The trends also exhibited directional asymmetry between north–south and east–west movements, indicating that a single fixed split is unlikely to be optimal for all corridor directions.

Offset experiments further examined corridor-level coordination effects. Average waiting time was analyzed across offset combinations between intersections. Certain offset settings produced noticeably lower corridor waiting times, whereas mismatched offsets amplified queue accumulation along the arterial. These results suggest that corridor efficiency depends on both local left-turn protection intensity and inter-intersection timing coordination.

5.5.2 Multi-agent policy evaluation

Learned corridor policies were compared with fixed-α baselines under deterministic traffic using the same evaluation metrics defined in Section 5. Table 4 summarizes average safety cost and average travel time (TT) across six fixed-timing configurations and the learned multi-agent PPO policy.

TABLE 4. Corridor-level PPO evaluation compared with fixed-α baselines.

Several observations emerge from Table 4. First, fixed-α control is highly fragile in the corridor setting: moderate increases in protected-turn allocation (α = 0.1 and α = 0.2) lead to severe efficiency degradation, with average travel time reaching 146.87 s and 212.56 s, respectively—three to five times higher than under the remaining configurations. This sensitivity underscores the difficulty of selecting a single timing split that remains effective under corridor-wide congestion propagation.

In contrast, the learned PPO policy achieves the best performance on both operational efficiency and surrogate safety among all evaluated controllers. PedSLight attains the lowest average travel time (42.52 s) and the lowest TTC-based safety cost (1.662), outperforming the strongest fixed-α baselines on each metric individually. Relative to the best fixed-timing configuration for travel time (α = 0.4, 43.07 s), PPO reduces average travel time by 1.3%; relative to the lowest-cost fixed configuration (α = 0.5, 1.698), it further reduces safety cost by 2.1%. More importantly, PPO is the only policy that simultaneously improves both metrics, whereas fixed-timing strategies exhibit a pronounced efficiency–safety trade-off (e.g., α = 0.5 lowers safety cost relative to α = 0.4 but increases travel time from 43.07 s to 48.55 s).

These results suggest that adaptive, cycle-level signal allocation is particularly valuable in multi-intersection corridors, where static timing parameters are difficult to tune and local control decisions interact through shared traffic dynamics. Although corridor training remains sensitive to reward scaling and multi-agent update stability, the evaluation demonstrates that the PedSLight formulation can be extended beyond a single intersection with measurable gains in both efficiency and safety.

6 Conclusion and discussion

This paper presented a safety-aware signal control framework for mixed vehicular–pedestrian intersections under dynamic traffic demand. A cycle-based reinforcement learning controller was developed by integrating a TTC-based near-conflict safety metric within a safety-aware weighted optimization framework to balance traffic efficiency and surrogate conflict exposure. Single-intersection experiments demonstrate that the proposed controller achieves a consistent efficiency–safety trade-off across varying traffic regimes, with more evident advantages under higher congestion levels. Compared with fixed-time and actuated baselines, the learned policy maintains competitive operational efficiency while mitigating surrogate conflict exposure. A three-intersection corridor extension further shows that the learned policy achieves the lowest estimated average travel time and TTC-based surrogate conflict cost among the evaluated controllers.

Future work will expand the simulation to larger-scale traffic networks and investigate cooperative multi-agent training, allowing neighboring signal controllers to share information and jointly optimize network-wide traffic efficiency and pedestrian safety.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this manuscript, the authors used ChatGPT solely for English-language polishing. All outputs were reviewed and revised by the authors, who take full responsibility for the content of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Konstantinos Koufos, Queen’s University Belfast, United Kingdom

Reviewed by: Mohammad Reza Jabbarpour, Niroo Research Institute, Iran

Mohammed Anis Oukebdane, Yıldız Technical University, Türkiye