3.1.1 Cloud-based global scheduling

Cloud-based global scheduling assigns decision authority to the cloud control center and aggregates cross-regional traffic information via cellular backhaul, thereby supporting macroscopic orchestration from a wide-area perspective. Typical tasks include global optimization problems such as cross-segment platoon routing, formation-density regulation, speed-limit control, and coordinated signal timing [24, 25]. The main advantage of this mode is that it avoids the local optimum induced by a single-segment perspective and enables system-level resource allocation across multiple platoons and intersections. The CARMA project is a representative application [16]. It aims to develop and test cooperative automated driving technologies based on a distributed control system, supported by ultra-low-delay and highly reliable cloud infrastructure with 5G access. In addition, the model-free multi-objective optimization controller proposed by Mahbub and Malikopoulos [26] performs centralized computation in the cloud and jointly optimizes safety margins, traffic efficiency, and energy consumption during platoon formation, thereby providing a globally optimal strategy for formation construction in mixed traffic. However, cloud-based scheduling faces three major constraints. First, the end-to-end delay of cellular networks is typically on the order of tens to hundreds of milliseconds, making it difficult to satisfy millisecond-level control-cycle requirements. Second, cross-regional information aggregation relies on base-station coverage and network-slicing quality, making data integrity difficult to guarantee in weak-coverage areas such as tunnels and mountainous regions. Third, the planning-execution mismatch caused by state desynchronization between the cloud and vehicles may be amplified by HDV disturbances. Therefore, cloud-based global scheduling is more suitable as a long-time-scale strategy generation layer than as a real-time control execution layer, and its effectiveness depends on co-design with vehicle-side local control.

From a deployment perspective, cloud-based global scheduling requires continuous wide-area cellular coverage, reliable uplink and downlink transmission of platoon states, unified cloud-edge data interfaces, and sufficient backhaul capacity. Since HDV states are usually obtained indirectly from onboard perception or roadside sensing, blind areas caused by tunnels, ramps, occlusions, or sparse RSU deployment should be compensated for by vehicle-side fallback perception and short-horizon prediction. Packet loss, delay jitter, and cellular handover may desynchronize cloud decisions and vehicle states; therefore, cloud commands should be limited to low-frequency tasks such as route guidance, speed advisory, formation management, and multi-platoon scheduling, while safety-critical spacing control and emergency braking should remain on the vehicle side. In large-scale road networks, computational scalability also becomes a major constraint, requiring regional decomposition, edge pre-processing, and hierarchical scheduling to avoid centralized overload as platoon number and HDV density increase.

3.1.2 Roadside edge coordination

Roadside edge coordination relocates the decision-making center to RSUs or edge computing nodes, where information aggregation, conflict coordination, and command issuance are performed within a regional scope. Compared with cloud-based scheduling, the physical proximity of roadside nodes reduces communication delay to the order of tens of milliseconds, making it better aligned with vehicle control cycles. Meanwhile, roadside multimodal perception systems enable blind-spot perception enhancement for occluded scenarios, non-line-of-sight targets, and distant HDVs, thereby compensating for the limited sensing capability of onboard perception [13]. In mixed traffic environments, roadside nodes fuse CAV state information obtained via V2I communication with HDV trajectories acquired through roadside perception, enabling local operations such as merging coordination, priority assignment, and platoon reconfiguration [27]. Gao et al. [28] proposed a coordination strategy between CAV platoons and RSUs to prevent potential collisions under abnormal traffic conditions. Zhang et al. [29] developed a centralized optimization framework for vehicle-signal coupling at intersections via roadside nodes, significantly improving traffic efficiency in mixed traffic. Zhao et al. [30] proposed a key-point matrix based on image processing and V2I technologies to manage infrastructure components associated with each key point.

However, the effectiveness of roadside coordination is constrained by three major factors. First, the coverage radius of a single RSU is limited, requiring state synchronization and authority handover among multiple RSUs for cross-regional platoons. Second, when roadside nodes fail, the platoon must revert to a fallback mode based on V2V communication or onboard perception, and inadequate switching mechanisms may lead to transient performance degradation. Third, roadside perception strongly depends on calibration accuracy and target association algorithms, and false positives and missed detections remain challenging in dense and highly interactive traffic scenarios. Overall, roadside edge coordination exhibits dual advantages in reduced communication delay and enhanced perception for medium-scale platoon control on structured roads. However, it should be co-designed with vehicle-side autonomous control to establish redundancy mechanisms and ensure system resilience.

Practical roadside-edge coordination depends strongly on RSU density, sensor placement, calibration accuracy, time synchronization, and edge computing capacity. To support stable platoon control, roadside sensors must provide sufficient longitudinal and lateral coverage for HDV detection, cut-in recognition, and conflict prediction. However, occlusion, adverse weather, dense traffic, and sensor misalignment may lead to missed detection or target association errors, which directly affect safety constraints used by downstream controllers. Communication reliability between vehicles and RSUs is also critical because packet loss or unstable V2I links may interrupt local coordination commands during merging, lane changing, or platoon reconfiguration. In addition, the edge node must process multi-target perception, trajectory prediction, and local optimization within a limited control window. When the number of vehicles increases, multi-RSU coordination, authority handover, and load balancing become necessary to maintain scalability across connected road segments.

3.2.1 Communication-topology coordination

Communication topology defines inter-vehicle communication links and directly determines information propagation paths, coordination efficiency, and system scalability. In pure CAV platoons, topologies such as predecessor-following (PF), bidirectional (BD), and predecessor-leader following (PLF) have been extensively studied. However, in mixed traffic, the random cut-in of HDVs may disrupt the predefined topology. When an HDV cuts in between two CAVs, the following vehicle loses its direct communication link with the preceding vehicle, and the topology degrades from PF to perception-based single-predecessor following, leading to increased information transmission delay and reduced disturbance suppression capability [31].

To address sudden topology changes, existing studies have evolved along three main directions. The multi-predecessor topology enables CAVs to obtain the states of multiple preceding vehicles via perception or multi-hop communication, thereby enhancing disturbance suppression capability [32, 33]. Wang et al. [34] proposed a bidirectional spacing balance strategy for CAV platoons, in which vehicles maintain a balanced position between the preceding and following vehicles using bidirectional multi-hop spacing information. Zhan et al. [35] partitioned a mixed vehicle platoon into multiple interrelated sub-platoons, where each CAV simultaneously acts as the tail vehicle of the preceding sub-platoon and the lead vehicle of the following sub-platoon. The virtual leader topology employs RSUs or cloud infrastructure to broadcast a virtual reference trajectory, allowing all CAVs to track a unified target rather than relying on inter-vehicle communication links, but it imposes stringent requirements on infrastructure communication quality [36]. Adaptive topology reconfiguration dynamically adjusts communication links following HDV insertion, such as splitting the platoon into multiple sub-platoons or establishing communication links that bypass the HDV. Nevertheless, transient performance degradation during topology switching and the design of switching conditions still lack rigorous theoretical guarantees [37, 38].

In terms of applicability, the multi-predecessor topology can significantly improve robustness when communication bandwidth permits. The virtual leader topology is well-suited for structured roads with mature infrastructure support, and adaptive topology reconfiguration demonstrates strong potential in scenarios with frequent topology changes. However, all three approaches require a careful trade-off between information redundancy and communication overhead.

3.2.2 Information-interaction coordination

The information-interaction mechanism determines when and what to communicate, directly affecting communication load, information timeliness, and coordination consistency. In mixed traffic, V2X communication faces bandwidth constraints and channel contention. As platoon size increases or the density of surrounding HDVs rises, high-frequency periodic broadcasting can lead to channel congestion, increased packet loss, and delay jitter, thereby degrading cooperative control performance. To alleviate communication pressure, the event-triggered mechanism triggers information transmission and control updates only when local state errors exceed predefined thresholds. Wakasa et al. [39] designed an event-triggered switching strategy for platoon merging and splitting scenarios, which balances communication savings and control accuracy by dynamically adjusting triggering thresholds. Chen et al. [40] proposed a hierarchical event-triggered algorithm to enable optimal control implementation when adjacent mixed platoons approach intersections. However, event-triggered mechanisms face an intrinsic trade-off. Excessively small thresholds lead to frequent triggering and limited communication savings. Conversely, overly large thresholds result in delayed control updates and insufficient responsiveness in fast transient scenarios, such as sudden HDV cut-ins. The self-triggered mechanism predicts the next triggering instant at each trigger, thereby avoiding the computational burden of continuous real-time error monitoring [41], but its prediction accuracy depends on system models and the estimation of disturbance bounds, limiting its reliability in highly uncertain environments. The asynchronous update mechanism allows different vehicles to update control actions at different frequencies and time instants, reducing reliance on global clock synchronization [42, 43]. However, under asynchronous conditions, more stringent conditions are required to guarantee consensus convergence, and the associated analysis complexity increases significantly.

From an engineering perspective, periodic communication remains a baseline solution in low-density scenarios with reliable communication quality. Event-triggered mechanisms are well-suited for medium-density scenarios with limited computational resources. Asynchronous update mechanisms offer a promising theoretical framework for high-density scenarios with dynamically changing topologies; however, several key challenges, such as triggering condition design, consensus verification, and anomaly handling, must still be addressed before large-scale real-world deployment.

Compared with infrastructure-supported architectures, distributed cooperative control reduces dependence on cloud or roadside facilities, but it imposes stronger requirements on vehicle-side communication, perception, and computation. Reliable deployment requires stable V2V neighbor discovery, robust topology reconfiguration protocols, and onboard fallback strategies when communication with the predecessor, leader, or adjacent CAVs is interrupted. As platoon size increases or HDVs frequently cut in, the number of communication links and local consistency constraints may grow rapidly, causing channel congestion and increasing the difficulty of maintaining string stability. Although computation is distributed among vehicles, the burden is not always uniform: CAVs adjacent to HDVs or topology-changing regions usually need to perform additional prediction, safety checking, and control adaptation. Therefore, practical distributed platoon control should jointly consider communication-load reduction, local computational load balancing, and fail-safe operation under topology degradation.

3.3.1 Functional hierarchical coordination

Functional hierarchical coordination divides the system into a planning layer and an execution layer according to task functions. The upper layer undertakes global or regional resource allocation, including trajectory planning and platoon formation scheduling, whereas the lower layer is responsible for trajectory tracking, local collision avoidance, and real-time control [46]. In mixed traffic scenarios, closed-loop information exchange between layers is achieved through well-defined input-output interfaces [47]. The upper layer can leverage cloud or roadside computing resources to address HDV behavior prediction and multi-vehicle coordination, generating macroscopic strategies that satisfy platoon integrity and efficiency objectives [48, 49]. The lower layer performs trajectory tracking based on onboard perception and V2V communication, and triggers safety protection mechanisms under abnormal conditions such as forced cut-in by HDVs and communication interruption [50]. Li et al. [51] proposed a two-layer control optimization scheme, in which the upper layer optimizes tracking performance, while the lower layer adopts a controller matching technique to guarantee the asymptotic stability of the vehicle platoon and achieve traffic wave dissipation. Ma et al. [52] proposed a hierarchical intersection management model. In this framework, the upper and middle layers handle signal optimization and phase scheduling, respectively. Meanwhile, the bottom layer implements CAV platoon control in the buffer zone and trajectory planning in the passing zone. Zhang and Du [53] proposed a hierarchical framework in which the upper layer generates platoon reference trajectories through virtual leader coordination, while the lower layer adopts model predictive control to achieve decoupled longitudinal and lateral tracking.

The key design issues of functional hierarchical coordination include the following. First, the representation of inter-layer outputs directly affects the degrees of freedom available to the lower-layer controller and the computational burden on the upper layer. Second, the bandwidth and delay of the feedback channel determine how rapidly the upper layer can perceive execution deviations, so excessively slow feedback may lead to planning-execution mismatch. Third, the constraint transfer mechanism must ensure the feasibility of upper-layer planning results subject to vehicle dynamics and safety constraints at the lower layer. Otherwise, inter-layer conflicts may arise. Overall, functional hierarchical coordination exhibits stable performance in scenarios with clear task division and predictable communication delay. However, under complex conditions involving rapidly changing HDV behavior and frequent topology switching, fixed inter-layer interface designs may become a performance bottleneck, requiring adaptive authority switching and emergency fallback mechanisms.

3.3.2 Spatiotemporal multi-scale coordination

Spatiotemporal multi-scale coordination decomposes the system along two dimensions, temporal update period and spatial coverage, thereby enabling appropriate allocation of computational load and tiered satisfaction of real-time requirements. In an integrated vehicle-road-cloud architecture for mixed traffic, the cloud, roadside, and onboard layers exhibit inherent differences in spatiotemporal scales. The cloud typically updates on the order of seconds to minutes and spans multiple road segments, making it suitable for macroscopic route optimization and cross-regional resource scheduling [54]. Roadside nodes update on the order of hundreds of milliseconds to seconds and cover a single intersection or road segment, making them responsible for local coordination and conflict detection [55]. Onboard control cycles are on the order of tens to hundreds of milliseconds, rely only on local neighborhood information, and focus on trajectory tracking and emergency avoidance [56]. This spatiotemporal separation decouples slow-timescale decision-making from fast-timescale control, preventing high-frequency control from being blocked by low-frequency tasks. Coppola et al. [57] designed a three-layer architecture for heterogeneous nonlinear electric vehicle platoons. In this architecture, the cloud generates energy-optimal speed profiles at second-level intervals, the roadside layer adjusts platoon spacing at hundred-millisecond-level intervals, and the onboard layer executes motor torque control at ten-millisecond-level intervals. Through time-scale separation, this architecture achieves coordinated optimization of energy consumption and safety. Yang et al. [58] proposed a hierarchical and deployable cooperative driving framework. At the vehicle level, a state transition graph was designed for different operating modes of CAVs. At the intersection level, a mixed-integer linear programming problem was formulated to optimize signal timing schemes and the arrival times of CAVs. At the corridor level, link performance functions were used to calculate the total delay of each coordinated phase at each intersection, and a linear programming problem was formulated to optimize the offset in each cycle, which was then transferred to the intersection level.

However, multi-scale coordination faces two types of coupling problems. First, in strong transient scenarios such as sudden HDV cut-in, the disturbance time scale approaches the control cycle, causing conventional scale separation to fail. Second, the delay and uncertainty of cross-scale information transmission may lead to state inconsistency, such that the traffic situational awareness used for upper-layer decision-making may already lag behind actual onboard conditions. Therefore, the effectiveness of spatiotemporal multi-scale coordination depends on appropriate delay budgeting, state prediction compensation, and cross-layer fault isolation mechanisms. At the design stage, the coupling intensity across different scales should be verified through simulation, and sufficient safety margins should be reserved for the fast-timescale layer to accommodate upper-layer decision deviations.

For real-world deployment, hybrid hierarchical architectures require clearly defined interfaces among the cloud, roadside edge, and vehicle-side controllers. Cross-layer communication must provide reliable timestamps, state synchronization, and quality-of-service guarantees; otherwise, inconsistencies between upper-layer planning and lower-layer execution may lead to infeasible or delayed control commands. Task allocation is another key constraint. HDV prediction, local conflict detection, and platoon reconfiguration can be assigned to the edge, whereas emergency braking and tracking control must remain onboard due to strict latency requirements. When multiple road segments or intersections are involved, scalability depends on regional decomposition, inter-RSU coordination, and smooth authority handover between control zones. Therefore, hybrid architectures offer strong engineering potential, but their deployment requires careful co-design of communication reliability, computational load allocation, cross-layer fallback, and safety redundancy.

In summary, Table 2 compares mixed-platoon control architectures in terms of physical deployment, decision mechanism, communication delay, computational load, and applicable conditions. The technical trend has shifted from single-layer centralized coordination toward edge-assisted, distributed, and vehicle-road-cloud hierarchical architectures, shown in Figure 3. This evolution reflects the need to balance global coordination with local real-time response under HDV disturbances. However, several unresolved conflicts remain. Centralized architectures improve global optimality but suffer from latency, scalability, and single-point failure risks. Distributed architectures enhance fault tolerance and adaptability to local cut-ins, but their global performance and string-stability guarantees become difficult to maintain under incomplete information. Hybrid hierarchical architectures provide a promising compromise, yet they introduce cross-layer inconsistency, authority allocation, and delay-budgeting problems. Therefore, future architectural design should move from static topology selection to adaptive architecture-control co-design, in which decision authority, communication topology, and fallback modes are dynamically adjusted according to traffic state, HDV uncertainty, and communication quality.

Figure 3 Control architecture of the mixed-platoon control system

From the perspective of control design, architecture selection is essentially a three-way trade-off among information availability, computational resource allocation, and fault-tolerant redundancy. It directly constrains the state variables available to downstream control laws, the solution time window, and the allowable uncertainty margin. The limiting effects of these three architectures on control performance should therefore be jointly incorporated at the controller design stage. Future research should leverage larger-scale road-test data and integrated vehicle-road-cloud experimental platforms to systematically validate different topological architectures in terms of cross-scenario stability, degree of communication dependence, and deployment cost. Meanwhile, adaptive and data-driven methods should be integrated to develop coordinated topology-control mechanisms that can adaptively reconfigure according to traffic states and communication conditions, thereby providing more engineering-deployable architectural choices for platoon control under mixed traffic conditions.

4.1.1 Heterogeneous car-following mechanisms

Heterogeneous car-following mechanism modeling is typically built upon classical car-following models, such as the intelligent driver model [60] and the optimal velocity model (OVM) [61]. In mixed traffic scenarios, it characterizes the longitudinal response characteristics of different driving styles by introducing parameter heterogeneity and structural heterogeneity. Parameter heterogeneity modeling treats key parameters, such as desired time headway and maximum acceleration, as driver-dependent random variables. Parameter distribution methods estimate parameter probability densities via population-level distribution fitting and compute confidence intervals of HDV trajectories [62]. Parameter clustering methods group drivers into discrete categories, such as conservative, normal, and aggressive drivers, assign each category a representative parameter set [63, 64], and yield category labels and corresponding deterministic parameter sets. Online identification methods update parameters in real time based on historical trajectories, producing time-varying parameters and their covariance [65]. Structural heterogeneity modeling explicitly incorporates human factors into the model equations. Reaction delay is introduced as fixed or time-varying delays to capture throttle-brake response lag [66]. Perception error is modeled by superimposing noise on leading-vehicle state observations to simulate inaccuracies in visual distance and speed estimation [67, 68]. Attention models capture drivers' selective attention mechanisms by introducing weighted combinations of leading-vehicle states or multi-predecessor information fusion schemes [53, 69].

Regarding output representation, heterogeneous car-following models typically provide point predictions, parameter intervals, or probability distributions of HDV longitudinal acceleration or position over a 1–3 s horizon, which can be leveraged by CAV controllers for feedforward compensation or robust constraint design. However, these models face three primary limitations. First, parameter identification requires an observation window of several to tens of seconds, hindering timely convergence in scenarios with frequent topology changes. Second, deterministic model structures are inadequate for capturing deeper factors such as drivers' cognitive blind spots, emotional fluctuations, and subjective preferences. Third, their extrapolation capability is limited, leading to a significant degradation in prediction reliability under extreme scenarios outside the training distribution.

4.1.2 Lane-changing and cut-in decision modeling

Lane-changing and cut-in decision modeling aims to characterize the triggering conditions, execution constraints, and sources of stochasticity in HDV lateral motion, thereby providing decision support for CAVs to proactively adjust inter-vehicle spacing or trigger yielding strategies. Existing methods mainly follow two modeling paradigms. Safe-gap-based rule models abstract lane-changing decisions into a set of explicit constraints. Specifically, longitudinal safety gap constraints require that the relative distance and velocity relative to target-lane vehicles satisfy minimum time headway or time-to-collision (TTC) thresholds [70]. Meanwhile, lateral execution constraints bound the peak lateral velocity and acceleration during the lane-change maneuver. The decision logic typically triggers a lane change when all constraints are satisfied simultaneously [71]. Such models feature simple structures and high computational efficiency, and can be directly formulated as safety constraints for CAVs. However, they fail to adequately capture drivers' subjective preferences and risk tolerance, making it difficult to explain variations in lane-changing timing across different driving styles. Utility-based rational choice models, in contrast, assume that drivers decide whether to change lanes by comparing the overall utility of the current and target lanes [72]. Deterministic utility models trigger lane changes when the utility difference exceeds a threshold, yielding binary decisions [73]. Random utility models treat the threshold as a random variable following a Gumbel distribution, yielding lane-changing probabilities [74, 75], which can be used in a Monte Carlo simulation to generate multiple possible trajectories. These models can capture drivers' trade-offs among different factors through parameter calibration, but the functional form of the utility function and the specification of weighting coefficients lack a unified standard, and both interpretability and prediction accuracy heavily depend on prior design.

In terms of uncertainty representation, lane-changing models typically output lane-changing probabilities, confidence intervals of triggering times, or feasible trajectory sets, enabling CAVs to allocate safety margins probabilistically or adopt conservative decisions under worst-case scenarios. However, most existing models are developed under regular driving assumptions and exhibit limited capability in capturing irrational behaviors such as forced lane changes and aggressive gap insertion, resulting in high prediction failure rates in high-density and high-conflict scenarios.

4.2.1 Short-term tajectory prediction

Short-term trajectory prediction aims to infer future sequences of position, velocity, and acceleration based on the historical motion states of HDVs and the surrounding traffic context. Existing methods mainly adopt three types of neural network architectures. Long Short-Term Memory (LSTM) networks capture temporal dependencies in trajectories through gating mechanisms. For highway ramp merging scenarios, Wang et al. [76] used LSTM to encode historical HDV trajectories and predict the speed and lateral position at the merging time, providing inputs for higher-level CAV decision-making. Liu et al. [77] used LSTM to infer HDV trajectories over the next 3 s and lane-changing intentions, enabling CAVs to initiate yielding 1–2 s in advance and attenuate cut-in disturbances. LSTM input features typically include the kinematic state of the target vehicle, the relative states of neighboring vehicles, and road geometry information. LSTM outputs can be formulated as either point predictions or distributional predictions, with the latter achieving multimodal representation through mixture density networks or conditional variational autoencoders [78]. Transformer architectures mine behavioral patterns over longer time windows through self-attention mechanisms [79]. Liu et al. [80] used a Transformer to predict lane-changing intention and trigger CAV yielding-assistance control, thereby conveying cooperative intentions to HDVs through explicit actions such as deceleration and improving interaction predictability. Transformers are advantageous in that they can process long sequences in parallel without manually specifying the attention range. However, their inference delay may approach the upper bound of typical control periods in long-sequence and multi-target scenarios. Graph Neural Networks (GNNs) model multi-vehicle interactions as spatiotemporal graphs, where nodes represent vehicle states and edges represent interaction relationships. Through message passing mechanisms, GNNs aggregate local neighborhood information and update node representations. GNNs can naturally handle scenarios with dynamically changing topologies and variable numbers of vehicles, but graph construction strategies have a significant impact on prediction performance, and unified design guidelines remain lacking.

In terms of performance evaluation, the average displacement error of short-term trajectory prediction within a 1–3 s horizon is typically on the order of 0.5–1.5 m, but the error accumulates exponentially as the prediction horizon increases. Multimodal prediction can cover more than 90% of real trajectories, but it requires a tradeoff between the number of modes and inference efficiency. These methods are suitable as upstream modules for CAV controllers, providing rich prior information that still requires downstream processing for model predictive control or robust optimization.

4.2.2 Driving intention recognition

Driving intention recognition infers the discrete decision-making states of HDVs at a higher level of abstraction, rather than continuous sequences of trajectory points. Its primary value lies in providing early warnings for CAVs. Unlike trajectory prediction that relies on explicit lateral displacement, intention recognition infers driver intentions 1–3 s prior to the actual maneuver [81]. Existing methods mainly adopt two modeling frameworks. Discrete classification models formulate intention recognition as a multi-class classification problem, taking kinematic, environmental, and interaction features within a fixed time window as inputs and outputting posterior probabilities for each intention class [76]. These methods usually achieve high recognition accuracy, but they are sensitive to class imbalance and distribution shift. Sequence labeling models formulate intention recognition as a temporal labeling task, model intention transition probabilities in the state space, and learn feature representations and transition patterns automatically. Lu et al. [82] proposed a joint prediction system combining long-term and short-term prediction algorithms to predict drivers' cut-in intentions into vehicle platoons. Guo et al. [83] used LSTM to implicitly learn traffic patterns and driver behavior, thereby estimating and predicting partially observable microscopic traffic states. Such methods can output the timing and duration of intention transitions, providing CAVs with finer-grained temporal information, but training requires complete sequence annotations and thus incurs high data costs.

In terms of feature design, intention recognition often incorporates multimodal information in addition to kinematic inputs. Specifically, in-vehicle sensors can increase the recognition lead time, whereas external interaction features enhance the distinguishability of lane-changing intentions. It should be emphasized that driving style and driving intention are fundamentally distinct. Driving style is a long-term characteristic that remains stable across scenarios and can be extracted through clustering or transfer learning; whereas driving intention is a short-term state associated with a specific scenario and must be inferred in real time. In terms of real-time performance, the inference delay of intention recognition is typically on the order of tens of milliseconds, satisfying control-cycle requirements. However, long observation windows reduce the response speed to sudden intentions, requiring a tradeoff between recognition reliability and recognition lead time.

4.3.1 Leader-follower games

Leader-follower games characterize hierarchical interaction processes by distinguishing between leaders and followers. In mixed traffic scenarios, this framework is commonly used to model the conditional responses of HDVs to CAV actions. One representative approach treats traffic signals or upper-level control authorities as leaders and platoons or individual vehicles as followers, focusing on the optimization of vehicle-signal coordination strategies. Zhang et al. [29] proposed an ID3QN framework that solves a Stackelberg game via multi-agent reinforcement learning, thereby improving overall intersection traffic efficiency. Another approach considers HDVs involved in potential conflicts as leaders and CAV platoons as followers. Huang et al. [84] resolved conflicts by analyzing variations in HDV payoffs under different CAV acceleration strategies. Li et al. [85] modeled HDV yielding or competitive merging decisions as a Stackelberg game, where CAVs influenced the HDV utility function by selecting merging timing and lateral velocity, and HDVs chose optimal responses based on their own safety constraints and time costs. Chen et al. [86] employed a Stackelberg game to characterize the asymmetric interaction between CAVs and HDVs, enabling CAVs to make safe decisions based on predicted optimal HDV responses.

In terms of outputs, leader-follower game models typically provide HDV reaction functions or response probabilities, which can be used by CAVs for scenario simulation, strategy evaluation, or counterfactual reasoning. A key issue of such models lies in the validity of leader assignment. In vehicle-signal coupling scenarios, signal control naturally serves as the upper-level decision-maker, making the leader designation reasonable. In vehicle-vehicle games, treating HDVs as followers aligns with the proactive coordination paradigm of CAVs. However, HDVs observe CAV intentions with delays and noise in practice, making the complete information assumption difficult to strictly satisfy. Furthermore, the presence of multiple equilibria increases prediction uncertainty. Since multiple optimal HDV responses may exist under the same CAV strategy, it is necessary to introduce equilibrium selection criteria or data-driven calibration approaches to identify the most likely outcome.

4.3.2 Cooperative interaction games

Cooperative interaction games emphasize improving overall payoffs through cooperation and focus on coalitional stability and payoff allocation mechanisms. In mixed traffic scenarios, this framework is used to characterize the interaction patterns of subgroups of cooperative HDVs. Existing methods mainly follow two modeling paradigms. Coalition formation mechanisms model HDV-CAV cooperation as coalition games, where each coalition corresponds to the cooperative driving state of a group of vehicles, and the coalitional value function quantifies the benefits of cooperation, such as fuel savings and travel-time reduction. Using solution concepts such as the core or the Shapley value, these methods identify HDVs with incentives to join CAV platoons and determine payoff allocation schemes. Zhu et al. [87] considered platoon formation in mixed traffic as a coalition game process, selected vehicles suitable for leadership through leader eligibility assessment, and developed an obstacle-aware coalition formation algorithm to obtain an approximately optimal coalition structure. The outputs of such methods include cooperative HDV sets, coalition stability indicators, and payoff allocation schemes, which can support the design of CAV platoon recruitment strategies. However, the combinatorial complexity of coalition formation grows exponentially with the number of vehicles, making real-time solutions still difficult in medium- and high-density traffic.

Nash bargaining mechanisms are primarily concerned with localized cooperation scenarios. Jin et al. [88] investigated the mixed traffic merging problem on variable-curvature roads by modeling lane-changing negotiation between CAVs and HDVs as a Nash bargaining problem. By constructing a joint utility function and solving for the bargaining solution, they predicted mutually beneficial strategies for both parties. The outputs of such methods include cooperative propensity, yielding probabilities, or negotiated trajectory pairs, which can be directly used for CAV decision planning. The advantage of Nash bargaining lies in its ability to explicitly characterize mutually beneficial conditions; however, it relies on strong assumptions regarding participant rationality, information transparency, and commitment to protocol adherence. In practice, the willingness of HDV drivers to cooperate is difficult to predict, and protocol execution lacks enforcement mechanisms, which may lead to deviations between model predictions and actual behavior. Overall, cooperative games provide a clear analytical framework for mutually beneficial coordination, but their direct application in open mixed traffic still requires stronger incentive design as well as protocol enforcement and assurance mechanisms.

Based on the above analysis, Table 3 compares HDV behavior modeling and prediction methods from the perspectives of modeling basis, input-output form, real-time performance, applicable scenarios, and limitations. The technical trend is moving from deterministic car-following and rule-based lane-changing models toward interaction-aware, multimodal, and uncertainty-aware prediction, as shown in Figure 4. Model-driven methods remain valuable because they provide interpretability and physically meaningful parameters, but they are limited in representing abrupt, irrational, or aggressive HDV behaviors. Data-driven methods improve nonlinear representation and multimodal prediction, yet their black-box nature, out-of-distribution sensitivity, and weak uncertainty calibration limit their direct use in safety-critical control. Game-theoretic models explicitly describe strategic interaction, but their rationality and complete-information assumptions are often inconsistent with real HDV behavior. A central unresolved conflict is that prediction accuracy alone does not guarantee control usefulness: trajectory prediction provides continuous constraints but may increase latency and accumulated error, whereas intention recognition offers earlier warnings but may suffer from misclassification risk. Future research should therefore develop control-oriented HDV prediction, where point, interval, probabilistic, and intention outputs are explicitly matched with deterministic MPC, robust MPC, stochastic MPC, or risk-aware control. More attention should also be paid to uncertainty calibration, online adaptation, and benchmark evaluation under cut-in, forced merging, and topology-changing scenarios.

Figure 4 Technical approach for HDV behavior modeling and prediction

From the perspective of control-oriented applicability, the output form of HDV behavior prediction directly determines whether downstream controllers can construct robust constraints. Specifically, point predictions naturally align with deterministic MPC, interval-based predictions facilitate robust MPC, and distributional predictions enable stochastic MPC and risk-aware optimization. Therefore, prediction methods should be selected based on control-oriented applicability rather than prediction accuracy alone. Future research should establish tighter integration pathways among the three categories of methods. On the one hand, the structural priors of model-driven methods and the rationality constraints of game-theoretic frameworks can be leveraged to improve the interpretability, sample efficiency, and generalization capability of data-driven models. On the other hand, data-driven approaches can be used to refine and extend the parameter distributions of mechanistic models and game-theoretic assumptions. Building on this basis, verifiable safety constraints and multi-scenario evaluation benchmarks should be introduced to systematically enhance the reliability and engineering deployability of mixed platoon control in real-world road environments.

5.1.1 Classical model-driven control

Classical model-driven control is built upon vehicle dynamics and HDV behavior models, and designs control laws through feedback regulation or receding-horizon optimization. In mixed platoons, representative approaches mainly include proportional-integral-derivative (PID) control, MPC, and adaptive control. PID control regulates tracking errors through proportional, integral, and derivative actions, exhibits relatively low dependence on model accuracy, and can serve as a baseline solution for longitudinal control [89]. Bandapally et al. [90] implemented CAV platoon control using PID and addressed HDV cut-in and crossing maneuvers by distinguishing between priority and non-priority external vehicles. MPC jointly optimizes multiple objectives, including safe spacing, fuel economy, and ride comfort, over a receding horizon, and enables anticipatory decision-making by predicting trajectories over the moving horizon [59, 91]. To cope with HDV uncertainty, stochastic MPC handles random disturbances through chance constraints [32, 50], whereas robust MPC accounts for worst-case scenarios using uncertainty sets [92]. For diverse scenarios such as curved-road driving [93], urban congested road segments [94], and urban intersections [95, 96], scenario-based MPC further customizes constraint sets for complex traffic environments [97]. Adaptive control compensates for model mismatch and external disturbances through online parameter estimation, allowing the controller to adapt to variations in HDV driving styles [98, 99]. These three classes of methods exhibit different tradeoffs between performance and complexity. PID incurs the lowest computational overhead, but typically produces larger tracking errors under strong disturbances. MPC achieves superior tracking accuracy and multi-objective optimization performance, but each optimization can require tens to hundreds of milliseconds [100]. Adaptive control has a computational burden between those of PID and MPC, but generally requires a parameter convergence window ranging from several seconds to tens of seconds. Fundamentally, classical model-driven control is limited by the tension between model mismatch and real-time performance, while the stochasticity and strong nonlinearity of HDV behavior are difficult to fully characterize using finite-dimensional analytical models. Therefore, such methods are currently more suitable for highway platooning scenarios with relatively stable traffic states and regular HDV driving styles. Incorporating data-driven identification and online model correction while preserving provable stability remains a critical research direction for extending these methods to complex mixed traffic.

5.1.2 Robust control

Robust control (RC) explicitly models system uncertainties during controller design by treating factors such as HDV behavior and communication delay as bounded disturbances or parameter perturbations, and accordingly constructs control laws that guarantee robust stability [101, 102]. Existing approaches can be broadly categorized into four classes. H_∞ robust control suppresses disturbance effects by minimizing the system H_∞ norm [103]. The Mousavi group has developed a systematic H_∞ control framework [104–107], including gain-scheduled H_∞ control to handle time-varying vehicle speeds and communication delay, output-feedback H_∞ control for disturbance attenuation and global stability, and mixed H₂/H_∞ control to balance robustness and energy optimization. Zhou et al. [108] designed an H_∞ controller that minimizes the worst-case damping ratio by considering the frequency characteristics of HDV acceleration disturbances. Min-Max MPC focuses on worst-case optimization. Chen et al. [109] proposed a leader-centered centralized Min-Max longitudinal control scheme, Hu et al. [110] drove the system toward desired states under worst-case disturbances within a disturbance set, and Li et al. [111] further developed a robust fuzzy MPC framework. Tube MPC mitigates disturbances in real time via error tube constraints [112]. Zhang et al. [113] employed tube MPC to ensure trajectory constraint feasibility, while Chen et al. [23] constrained the actual state within an error tube centered on the nominal trajectory and guaranteed asymptotic convergence. Safety-function-based RC provides formal safety guarantees using control barrier functions (CBF) [114], or reformulates time-varying HDV behavior into linear matrix inequality (LMI) problems for tractable solutions [115, 116]. These four classes exhibit different tradeoffs between conservatism and performance. H_∞ control offers high computational efficiency and guarantees worst-case performance, but requires predefined disturbance bounds, often leading to conservative control inputs and reduced fuel economy under nominal conditions. Min-Max MPC achieves moderate conservatism but incurs high computational cost and depends on convexity assumptions for convergence. Tube MPC reduces computational complexity by decoupling nominal trajectories and error tubes, although the tube width increases linearly with uncertainty bounds. Safety-function-based methods provide formal guarantees, but the intrinsic conservatism of LMI-based approaches may degrade performance.

The fundamental limitation of RC lies in the fact that simple bounded-disturbance models cannot adequately capture the heavy-tailed distribution and strong asymmetry of HDV behavior. Overly tight uncertainty bounds may lead to constraint violations, whereas overly loose bounds result in excessive conservatism. Therefore, RC is more suitable for safety-critical scenarios with relatively well-defined uncertainty bounds and tolerable conservatism. Future research necessitates establishing a tighter integration between data-driven uncertainty modeling and provable robustness.

5.2.1 Data-driven predictive control

Data-driven predictive control (DDPC) uses historical input-output data to construct prediction models, thereby replacing the mechanistic models of vehicle dynamics and HDV behavior used in conventional MPC. According to the data modeling strategy, DDPC methods can be divided into three categories. Reachable-set-based DDPC characterizes system uncertainty using reachable sets represented by matrix zonotopes. Based on this representation, Lan et al. [117] designed a robust MPC method that requires neither explicit HDV models nor delay parameters and can handle measurement noise. Although an effective reachable set can be constructed from only hundreds of trajectories, the method is relatively conservative. Koopman-operator-based DDPC transforms nonlinear mixed platoon systems into high-dimensional linear systems through lifting-based linearization [118]. Li et al. [119] combined Koopman-based extended dynamic mode decomposition with reachable sets and proposed a robust nonlinear DDPC method. This class of methods offers high online optimization efficiency, but offline training requires thousands to tens of thousands of high-quality trajectories. Willems-lemma-based DDPC is tailored to scenarios involving the coexistence of CAVs and HDVs. Wang et al. [120] proposed DeeP-LCC to realize model-free predictive control, and Li et al. [121] further developed RDeeP-LCC to enhance robustness. This class of methods requires maintaining high-dimensional Hankel matrices, which imposes a heavy burden on online optimization. Furthermore, when deep neural networks are used as prediction models, their coupling with MPC leads to nonconvex optimization problems, and the solution time may be several times longer than the control period. Overall, DDPC still faces challenges, including data scarcity in long-tail extreme scenarios, insufficient model interpretability, and difficulty in ensuring real-time online optimization. At present, DDPC is more suitable as a complementary module for model-driven control or as an offline benchmarking tool. Future research should further investigate the interpretability of black-box models, the efficient utilization of structured data, and approximate solution algorithms.

5.2.2 Reinforcement learning control

Reinforcement learning (RL) control treats CAVs as agents that interact with the environment. By designing the state space, action space, and reward function, RL enables agents to balance platoon stability with responsiveness to uncertain HDV behaviors through trial-and-error learning [122]. Existing RL-based control architectures can be broadly categorized into three classes. Multi-agent reinforcement learning (MARL) relies on independent decision-making by multiple agents and local information exchange, making it suitable for mixed platoon scenarios. Li et al. [123] proposed a dense-communication RL framework that mitigates traffic oscillations by enhancing inter-vehicle information exchange. Shi et al. [124] developed a longitudinal control strategy based on distributed deep reinforcement learning (DRL), embedding stochasticity in car-following behavior into the training environment. Wang et al. [125] designed a multi-agent DRL controller to address communication delay. Shi et al. [126] proposed MARL-based methods for platoon control and platoon formation. For the coordination of heterogeneous vehicle fleets, Gao et al. [127] proposed a hub-and-spoke graph RL framework for spatiotemporal cooperative decision-making, while Dai et al. [128] constructed a cooperative multi-agent DRL system. MARL methods can handle local observations and communication constraints, but they commonly suffer from low sample efficiency and unstable training [129]. Safe reinforcement learning (SRL) has attracted considerable attention because it can directly address safety constraints such as collision avoidance and formation keeping. Pan et al. [130] proposed a safety-supervised decentralized Proximal Policy Optimization (PPO) algorithm, which significantly improves fuel economy while ensuring safety and stability. Zhou et al. [131] embedded CBF into DRL as a differentiable network layer to provide system-level safety assurance. Yang et al. [132] integrated CBF into agents as a differentiable quadratic programming layer, mapping RL outputs into actions that satisfy safety constraints and enabling optimal speed advisory during green phases at signalized intersections. SRL provides formal safety guarantees through constraints such as CBFs, but it substantially narrows the exploration space, making the learned policies conservative. Moreover, designing safety constraints remains challenging in complex mixed traffic scenarios. Robust reinforcement learning (RRL) aims to improve policy adaptability to model uncertainty and anomalous behaviors. Bejarbaneh et al. [133] proposed an RRL framework that employs distributed observers and hybrid controllers to address uncertainties in heterogeneous vehicle dynamics. Chen et al. [134] proposed guided deep deterministic policy gradient (GuidDDPG), which combines the advantages of RL and conventional adaptive cruise control (ACC) through a goal-oriented reward function, thereby guiding the policy to converge toward safer and more efficient acceleration and deceleration patterns. RRL has made progress in reducing training costs and enhancing deployment practicality, but it still needs to trade off model generalization against computational efficiency.

These three classes of RL-based control architectures differ in their performance emphases. MARL enables distributed cooperative decision-making, but has relatively low sample efficiency and is sensitive to communication quality. SRL provides safety guarantees but restricts the exploration space, resulting in conservative policies. Robust and hybrid methods balance safety and deployment practicality, but their robustness remains insufficient under extreme anomalous scenarios. Overall, RL-based control still faces challenges, including high trial-and-error costs, difficulty in obtaining formal safety certification for black-box policies, and insufficient validation in real-world mixed traffic. At present, it is more suitable for offline policy mining and prior policy generation, or for hybrid use with model-driven methods to improve engineering deployability.

5.2.3 Data-driven adaptive control

Data-driven adaptive control replaces explicit mechanistic models with real-time operational data and constructs adaptive laws through online learning algorithms, enabling controllers to dynamically adjust their parameters in response to changes in HDV driving styles and evolving traffic states. Existing studies can be broadly divided into two categories. Neural-network adaptive control uses neural networks to approximate the complex nonlinear and time-varying characteristics of mixed traffic, among which Radial Basis Function Neural Networks (RBFNNs) are widely adopted because of their simple structure and strong approximation capability [135]. Dong et al. [116] addressed mixed-order nonlinear heterogeneous platoons by designing an RBFNN-based adaptive law to simultaneously approximate uncertain and nonlinear terms. Adaptive optimal control enables each CAV to independently adjust its control parameters based on local information within an optimization framework and to achieve near-global optimality through interactions with neighboring vehicles [136]. For example, Lan et al. [137] developed a data-driven control policy learning method based on adaptive dynamic programming. Jiang et al. [138] proposed a random-search-based adaptive policy learning method for mixed platoons at signalized intersections.

These two classes of methods exhibit a tradeoff between convergence speed and stability. Neural-network adaptive control has strong approximation capability and can handle pronounced nonlinearities, but its convergence may require tens of seconds to several minutes and is sensitive to measurement noise. Adaptive optimal control converges relatively faster and offers a degree of global optimality, but it is relatively dependent on V2V communication quality. When HDV behavior changes rapidly, adaptive laws often lag behind disturbances. To guarantee Lyapunov stability, the learning rate usually needs to be reduced, which further slows convergence. Therefore, data-driven adaptive control is currently more suitable for scenarios with slowly varying disturbances and stable communication conditions. Future research should establish systematic design principles that balance fast adaptation, noise suppression, and performance guarantees under safety constraints, while exploring coordination mechanisms with model-driven control.

5.3.1 Embedded fusion control

Embedded fusion control uses physical models to describe deterministic system dynamics and data-driven models to approximate uncertain or unmodeled components that are difficult to characterize accurately. In other words, it embeds one mechanism into a dominant algorithmic framework as a module or constraint. According to the fusion mode, embedded fusion control can be divided into three categories. Model-enhanced RL explicitly incorporates vehicle dynamics and safety constraints into agent training to improve physical consistency. For example, the DDPG-OVM method proposed by Qi et al. [139] design a deep reinforcement learning controller by using the optimal velocity model as a prior. Data-enhanced MPC embeds data-driven prediction into the MPC framework. Hu et al. [140] combined vehicle dynamics equations with an empirical database of queue dissipation, and estimated terminal conditions and energy consumption bounds through receding-horizon optimization. Long et al. [141] proposed a predictive control method based on physics-enhanced residual learning, which combines MPC with deep learning techniques to accurately predict the behavior of the preceding vehicle. Koopman-operator-based hybrid control maps nonlinear mixed platoon systems into a high-dimensional linear space through lifting-based linearization. Lyu et al. [142] proposed KoopLCC, which uses deep Koopman networks to extract driving styles and perform modeling in a high-dimensional linear space. They further developed a deep variational Koopman network, DVKoN [143], and combined it with a variational autoencoder to form a two-layer robust predictive controller, DVKoRPC.

These three categories exhibit different tradeoffs among physical consistency, data requirements, and training complexity. Model-enhanced RL reduces data requirements compared with pure RL by introducing strong physical priors, but physical constraints may impede gradient propagation through the policy network and prolong training. Data-enhanced MPC can flexibly adjust the relative weights of model-based and data-driven components and achieves relatively high online optimization efficiency, but it is more sensitive to the accuracy of the underlying model. Koopman-based control methods offer high online efficiency after training, but lifting and network training substantially increase offline computational complexity. Overall, embedded fusion control faces the challenge of quantitatively balancing the strength of physical priors, and physical constraints may render the optimization problem nonconvex. Therefore, embedded fusion control is more suitable for subsystems with well-understood physical mechanisms and certain requirements for safety and interpretability.

5.3.2 Hierarchical fusion control

Hierarchical fusion control constructs a multi-layer control architecture in which model-driven and data-driven methods are deployed at different layers. The output of an upstream algorithm serves as the input to a downstream algorithm, and each layer is dedicated to specific time scales or functionalities. According to the architectural design, it can be broadly divided into two categories. The decision-planning-execution three-layer architecture typically employs data-driven methods at the decision-making layer for high-level decision-making, model-driven methods at the planning layer to generate reference trajectories, and classical control at the execution layer for trajectory tracking [144, 145]. This structure enables clear task decomposition. Typically, the decision layer operates on a timescale of seconds, whereas the execution layer functions within tens of milliseconds. However, inter-layer coordination and time-scale alignment remain challenging, placing stringent requirements on the rapid response capability of lower layers. The feedforward-feedback two-layer architecture integrates data-driven predictive compensation with model-driven error correction [146]. The feedforward layer predicts future HDV trajectories based on historical data and generates anticipatory control inputs, while the feedback layer compensates for real-time tracking errors based on vehicle dynamics models. This architecture is simpler to coordinate and offers superior real-time performance, but it relies more heavily on the accuracy of feedforward prediction and the robustness of the feedback model.

Overall, the core challenge of hierarchical fusion control lies in the tension between inter-layer coupling and the hierarchical decoupling assumption. In mixed traffic, abrupt HDV behaviors tend to diminish time-scale separation across layers, thereby challenging the conventional fast-slow separation paradigm. In addition, global objectives may be partially localized during inter-layer propagation, which can degrade overall system performance. Therefore, hierarchical fusion control is more suitable for scenarios with well-defined hierarchical structures and clear time-scale separations. Future research requires the development of systematic design principles to resolve inter-layer conflicts and ensure objective consistency.

5.3.3 Joint optimization fusion control

Joint optimization fusion control, in contrast to explicitly layered hierarchical fusion control, integrates model-driven and data-driven modules within a unified optimization framework. Rather than forming a simple cascaded input-output relationship, the constituent algorithms are tightly coupled through shared objective functions or constraint sets. According to the fusion strategy, it can be broadly categorized into three classes. The integration of MPC and RL leverages MPC to solve constrained optimization problems over a short prediction horizon to ensure short-term performance, while RL learns policy parameters or objective weights over longer time scales, enabling the receding-horizon objective function to adapt to long-term cumulative returns [147]. This approach allows RL to optimize control parameters under the safety guarantees provided by MPC, but inconsistencies between their objective functions may arise, necessitating careful design of a unified reward function. Physics-data-driven joint modeling embeds the ego-vehicle dynamics model together with data-driven models of preceding and surrounding HDVs into a single optimization problem, where they jointly function as constraints or objectives [148]. This approach tightly couples model-driven and data-driven components, but incurs high computational complexity. Moreover, the nonconvexity of data-driven models may undermine the convex structure of MPC, making global optimality difficult to guarantee. The integration of uncertainty quantification and robust optimization incorporates uncertainty estimates obtained from data-driven models into a robust optimization framework, enabling explicit consideration of uncertainty effects on top of physical models. For example, Zhang et al. [149] proposed a coordinated control strategy that combines offline robust MPC with online data-driven control input mapping. This approach achieves a balance between coupling strength and computational complexity, but is highly sensitive to the accuracy of uncertainty representation.

Overall, as the degree of integration increases, the computational burden of joint optimization fusion control grows superlinearly, significantly exceeding that of hierarchical fusion approaches. Therefore, it is currently more suitable as a tool for theoretical exploration. Future research necessitates leveraging structured learning, distributed optimization, and approximate verification techniques to alleviate computational burden and enhance formal guarantees.

Overall, Table 4 summarizes the application scenarios and quantitative performance of different control methods, while Figure 5 illustrates their general control structure. The technical trend has evolved from analytically tractable model-driven control toward learning-enhanced and hybrid model-data-driven control. Model-driven methods provide explicit stability, safety, or robustness guarantees, but their performance depends on model fidelity and predefined uncertainty bounds. Data-driven methods, especially reinforcement learning and adaptive learning, improve adaptability to nonlinear HDV behavior and complex mixed-traffic interactions, but they still face weak interpretability, limited formal certification, and sim-to-real generalization problems. Hybrid model-data-driven methods attempt to combine physical constraints with learned prediction or policy modules, making them a promising direction for deployable mixed-platoon control. However, deeper integration also brings new conflicts: safety guarantees may be weakened by nonconvex learning modules, robust constraints may become overly conservative, and joint optimization may exceed real-time computational limits. Future research should focus on safety-certified learning control, uncertainty-aware robust optimization, lightweight distributed computation, and hardware-in-the-loop or field-test validation, so that adaptability, real-time feasibility, and formal control guarantees can be jointly achieved.

Figure 5 Algorithm-driven mixed platoon control architecture

Future research should focus on systematically validating existing methods in scenarios that closely reflect real-world constraints, developing a unified, interpretable, and provably safe fusion control framework, and designing lightweight control strategies with capabilities for online learning, adaptive weight adjustment, and cross-scenario generalization. These advances are critical to ensuring the robust performance of mixed platoons under out-of-distribution (OOD) conditions and complex traffic disturbances, thereby providing reliable support for engineering deployment.