Support Driving (SD) in Autonomous Systems

• Support Driving (SD) is a family of techniques that integrates human-in-the-loop control, remote assistance, and standard-definition map priors to improve safety and operational robustness.
• It employs methods such as haptic shared control, cooperative driver adaptation, and remote driving support, tuned for varying operational design domains and dynamic conditions.
• SD enhances system performance through measurable gains like reduced trajectory errors, improved localization accuracy, and effective traffic-flow stabilization using adaptive support strategies.

In recent autonomous- and assisted-driving literature, Support Driving (SD) is used in several closely related senses rather than as a single standardized subsystem. Across the cited works, it encompasses shared and cooperative driver assistance, remote driving support and intervention, safety-centric evaluation and perception, traffic-flow stabilization by support vehicles, and the use of scalable priors—especially standard-definition (SD) maps—to support localization, scene understanding, and trajectory prediction. Taken together, these strands frame SD as a family of methods that augment driving with external guidance, structured priors, or supervisory control in order to improve safety, controllability, efficiency, and robustness under operational constraints (Wada, 2019, Hans et al., 18 Jul 2025, Lucas-Estañ et al., 16 Jun 2026, Suzuki et al., 5 Aug 2025, Zhong et al., 6 Jan 2026).

1. Conceptual scope and system taxonomy

A central distinction in the SD literature is between direct operational support and indirect advisory support. In shared control, the human and machine jointly perform the same operational task, often through a haptic interface; cooperative control includes shared control but also broader collaboration across strategic, tactical, and operational layers. In high-level ADS support, the analogous distinction is between a Remote Driving System (RDS), in which the remote operator takes direct and full dynamic control, and a Remote Assistance System (RAS), in which the operator gives indirect guidance while the ADS maintains physical control and full dynamic driving task execution (Wada, 2019, Hans et al., 18 Jul 2025).

The selection of an SD architecture is explicitly tied to Operational Design Domain (ODD) analysis in the remote-support literature. A structured approach uses the PEGASUS six-layer model—Road Network Layout and Road types, Traffic Infrastructure, Temporary Modifications/Changes, Dynamic Objects, Environmental Conditions, and Communication Network/Digital Infrastructure—to determine where ADS capability gaps arise and whether those gaps require direct takeover or advisory support. The cited analysis concludes that neither RDS nor RAS is universally superior; suitability is layer- and use-case-dependent, with RDS favored in ambiguous, dynamic, or rapidly changing situations and RAS favored when strategic guidance is sufficient or digital infrastructure is constrained (Hans et al., 18 Jul 2025).

A second terminological axis concerns the frequent use of the initials “SD” for standard-definition maps. In that literature, SD maps are not themselves support driving, but they function as support priors for perception, topology reasoning, localization, and route-aware prediction. This suggests that recent SD research increasingly combines human-support paradigms with map-supported autonomy, making the term span both human-in-the-loop assistance and machine-inferred support from lightweight infrastructure priors (Zhang et al., 2024, Pei et al., 18 May 2025, Voloshyn et al., 1 Jul 2026).

2. Shared control, cooperative control, and driver adaptation

The most explicit human-centered formulation of SD appears in work on shared and cooperative control. Its motivating claim is that advanced driver assistance systems reduce workload and increase safety, but excessive use can impede skill development. The proposed remedy is task difficulty adjustment (TDA): adaptive assistance tuned to the driver’s current skill so that workload remains compatible with learning. The methodology is organized around three mechanisms—training in parts, simplifying, and guiding—and is instantiated through haptic shared control (HSC), where the driver and driver assistance system act jointly through steering-wheel torque feedback (Wada, 2019).

In the reverse-parking HSC experiment, assistance torque is modeled as

$$T_{das}(t) = -C_s|e(t)|\{\theta(t) - \theta^*_a(t)\},$$

where $e(t)$ is position error from the desired trajectory, $\theta(t)$ is current steering angle, $\theta^*_a(t)$ is the desired steering angle, and $C_s$ is the assistance gain. Three gain conditions were evaluated: $C_s = 0$, $0.5$, and $1.0$. Moderate and strong assist significantly reduced trajectory error and steering workload during assistance, and the resulting performance gains persisted after assistance was removed. The reported correlation between performance improvement during assist and post-assist skill improvement was $r = 0.730$ with $p = 0.007$, supporting the claim that properly tuned HSC can simultaneously reduce workload and improve skill (Wada, 2019).

A complementary line of work addresses driver adaptation through rapid style recognition. The k-means clustering-based support vector machine (kMC-SVM) method classifies curve-negotiating behavior into aggressive and moderate types using vehicle speed $e(t)$0 and throttle opening $e(t)$1 as the feature vector $e(t)$2. k-means is applied within each class to reduce data volume and overlap, after which an SVM with Gaussian kernel performs classification. Leave-1-out cross-validation is used, and for a large training set of $e(t)$3, kMC-SVM required about $e(t)$4 seconds whereas conventional SVM required over $e(t)$5 seconds. Reported online accuracies included $e(t)$6 and $e(t)$7 for one aggressive/moderate pair and $e(t)$8 and $e(t)$9 for another (Wang et al., 2016). A plausible implication is that adaptive SD can be personalized by coupling assistance authority to online estimates of driver style or control tendency.

3. Remote assistance, remote driving, and teleoperation infrastructure

Remote-driving SD research treats the in-vehicle Safety Driver (SD) as a fallback layer that compensates for Remote Driver (RD) performance limitations. In a real-world urban ODD study in Las Vegas, more than $\theta(t)$0 km of remote-driving data from $\theta(t)$1 RDs yielded $\theta(t)$2 disengagements attributable to human RD performance rather than technical failures or third-party misbehavior. The number of SD interventions per $\theta(t)$3 km decreases significantly within the first $\theta(t)$4 km of RD experience, with a marked learning plateau after roughly $\theta(t)$5–$\theta(t)$6 km. The common causes identified across the studies are braking too late for signs, traffic light went red, leaving the lane, and being impatient for other traffic or obstacles; the low-experience group dominates these event categories (Hans et al., 31 Mar 2025, Hans et al., 12 Mar 2025).

The remote-driving literature also formalizes intervention severity. One study introduces a classification spanning SD-S0 to SD-S3 for hazard and injury potential, SD-M0 to SD-M2 for traffic impact, and SD-U1 to SD-U3 for user- or scenario-specific effects. It further links intervention likelihood to harsh braking, harsh acceleration, and harsh steering events, with thresholds such as $\theta(t)$7, $\theta(t)$8, and $\theta(t)$9. Subjective RD feedback attributes many early difficulties to latency and the replacement of natural haptic feedback with visual feedback, reinforcing the view that SD in remote driving is a coupled human–interface–network problem rather than a purely behavioral one (Hans et al., 12 Mar 2025).

At the communication layer, teleoperated driving (ToD) imposes explicit uplink-heavy requirements. The cited 5G study reports $\theta^*_a(t)$0 Mbps uplink video and $\theta^*_a(t)$1 Kbps downlink commands per vehicle, with a target scalability of $\theta^*_a(t)$2 vehicles per $\theta^*_a(t)$3. MEC- or edge-based architectures outperform centralized ones: for $\theta^*_a(t)$4 ms and $\theta^*_a(t)$5 MHz, only $\theta^*_a(t)$6 of packets in a centralized architecture meet the $\theta^*_a(t)$7 ms uplink latency bound, versus $\theta^*_a(t)$8 for MEC@CN. The $\theta^*_a(t)$9 uplink latency percentile is about $C_s$0 ms lower for MEC@CN than for centralized deployment, and centralized architectures fail the downlink latency target while MEC-based architectures meet it consistently. The same study shows that duplexing and control-channel configuration matter: TDD1 (DDDSU), balanced TDD3, and FDD are better suited than TDD2, and the SC2 control-channel configuration improves the percentage of packets meeting the $C_s$1 ms requirement under larger video-processing delays (Lucas-Estañ et al., 16 Jun 2026).

4. Safety-critical evaluation, scenario synthesis, and accident understanding

One SD research direction addresses the scarcity of safety-critical data by synthetic scenario generation. For follow-up driving, a scenario-generation method models a leader and follower with kinematic equations and stochastic driver reaction times sampled from a gamma distribution with mean approximately $C_s$2 s, standard deviation approximately $C_s$3 s, minimum $C_s$4 s, and maximum $C_s$5 s. The simulation uses $C_s$6, vehicle length $C_s$7, and normally distributed initial positions, speeds, and accelerations. Its core safety indicator is Difference Space Stopping (DSS):

$C_s$8

Here, $C_s$9 denotes sufficient stopping space and $C_s = 0$0 denotes a safety-critical state. This provides an objective labeling rule for large synthetic follow-up-drive corpora used in validation of driver-assistance and autonomous functions (Schick et al., 2024).

A second strand critiques conventional route-completion metrics by introducing safety-centric closed-loop evaluation. Safe2Drive (S2D) adds $C_s = 0$1 common but challenging scenarios to Bench2Drive, focusing on work zones, pedestrian jaywalking, and occluded vulnerable road users. It defines SafeDriving Score (SDS) as

$C_s = 0$2

with the scenario-set score

$C_s = 0$3

The benchmark shows large degradations for two state-of-the-art end-to-end policies: LEAD drops from $C_s = 0$4 DS on Bench2Drive to $C_s = 0$5 DS on S2D and achieves SDS $C_s = 0$6, while SimLingo drops from $C_s = 0$7 DS to $C_s = 0$8 DS and achieves SDS $C_s = 0$9. Reported failure modes include poor work-zone understanding, red-light violations, and late or absent braking for pedestrians (Sahu et al., 29 May 2026).

Safe-driving perception is extended further by abductive accident understanding. The MM-AU dataset contains $0.5$0 in-the-wild ego-view accident videos, more than $0.5$1 million object boxes, $0.5$2 pairs of video-based accident reasons, and $0.5$3 accident categories. On top of it, AdVersa-SD combines an abductive CLIP model with Object-Centric Accident Video Diffusion (OAVD) to align normal, near-accident, and accident segments with accident reason, prevention advice, and category text. Reported diffusion results show CLIP$0.5$4 and FVD $0.5$5 for OAVD with abductive CLIP and bounding boxes, compared with CLIP$0.5$6 and FVD $0.5$7 for a vanilla CLIP baseline; qualitative results emphasize object-centric cause–effect reasoning and counterfactual prevention generation (Fang et al., 2024).

5. Standard-definition maps as support priors for perception, localization, and prediction

A major current use of support priors comes from standard-definition map integration. In online road-network perception and reasoning, rasterized SD maps and graph-based SD maps are fused with BEV encoders. On OpenLane-V2, adding SD maps to MapTR enables convergence in $0.5$8 epochs versus $0.5$9 for image-only models and yields a $1.0$0 relative improvement in mean average precision for online centerline perception, with MapTR+SD reaching $1.0$1 mAP versus a $1.0$2 baseline. For joint perception and reasoning, TopoNet+OSMR improves OLS from $1.0$3 to $1.0$4, while TopoNet+OSMG reaches $1.0$5 with only $1.0$6M parameters, compared with $1.0$7M for raster-based integration. Robustness analysis reports less than $1.0$8 performance drop under moderate SD-map translation error $1.0$9 m and rotation error $r = 0.730$0 (Zhang et al., 2024).

SEPT extends this line by combining rasterized and vectorized SD map representations through a hybrid feature-fusion pipeline. Its Feature Transformation module addresses map–BEV misalignment, its vector branch uses cross-attention for geometric and topological alignment, and its Dual Gated Feature Fusion module combines the branches. SEPT also introduces an auxiliary intersection-aware keypoint detection task. On OpenLane-V2, the reported OLS improves from $r = 0.730$1 for the TopoNet baseline to $r = 0.730$2 for SMERF and $r = 0.730$3 for SEPT, while $r = 0.730$4 rises from $r = 0.730$5 to $r = 0.730$6 to $r = 0.730$7 (Pei et al., 18 May 2025).

For localization, HOLO formulates fine-grained visual localization between multi-view images and SD maps as homography-guided pose estimation. It constructs image–map pairs that satisfy a planar homography constraint by projecting ground-view features into the BEV domain and enforcing semantic alignment with map features, then uses homography relationships both to guide feature fusion and to restrict pose outputs to a valid feasible region. The paper states that this is the first work to unify BEV semantic reasoning with homography learning for image-to-map localization and reports significant outperformance over existing state-of-the-art methods on nuScenes, while also supporting cross-resolution inputs through explicit homography modeling (Zhong et al., 6 Jan 2026).

Bayesian localization and lane-level map creation have also been built on SD maps. A simultaneous localization and multi-lane tracking filter uses GNSS, visual odometry, lane-marking edge detections, and an SD map to estimate vehicle 6D pose, map-relative position, and 3D traffic-line geometry. Traffic lines are represented as quadratic B-spline trajectories and estimated with a trajectory Poisson multi-Bernoulli mixture filter. In highway experiments, the reported localization RMS errors are $r = 0.730$8 m in $r = 0.730$9, $p = 0.007$0 m in $p = 0.007$1, and $p = 0.007$2 m in $p = 0.007$3, yielding an approximately $p = 0.007$4 improvement over a GNSS plus visual-odometry baseline; the resulting lane estimates align with satellite imagery up to some lateral offsets (Xia et al., 2024).

Support priors now also reach prediction. SD-RouteFusion uses a front-facing camera, vehicle kinematics, and a route derived from an SD map in a dual-hypothesis design with a gated classifier. On $p = 0.007$5 real-world driving scenarios across $p = 0.007$6 European countries and the U.S., incorporating SD-map routes yields a $p = 0.007$7 ADE improvement over an image-and-kinematics baseline, while the full fusion method achieves a $p = 0.007$8 ADE reduction over that baseline at an $p = 0.007$9-second horizon. The same work releases an SD-route generation toolkit for datasets containing ego pose and future trajectories (Voloshyn et al., 1 Jul 2026).

Map enhancement itself is becoming an SD function. SD++ uses road manuals and LLMs to enrich SD maps with lane widths, lane counts, bike lanes, and related attributes, without LiDAR, camera data, or manual annotation. Its knowledge-based algorithmic generation pipeline uses retrieval-augmented generation and deterministic map generation. On Argoverse 2 in Palo Alto and Miami, GPT-4o with one-shot generation and an improved prompt produces $e(t)$00 valid maps, Chamfer$e(t)$01 m, Chamfer$e(t)$02 m, and recall $e(t)$03; the same framework is demonstrated on both California and Japan (Diwanji et al., 4 Feb 2025).

A related support representation is online situation-aware drivable space estimation. SDS++ replaces the earlier SDS graph output with Artificial Potential Fields, introduces a graph-SLAM factor for line features based on bi-cubic implicit polynomial equations, models objects and lanes with sigmoid-plus-implicit APFs, and integrates domain knowledge to update drivable space according to the current situation. It is validated with raw vehicle data and an MPC-based planner, and is reported to improve planning robustness against localization noise while enabling trajectories that adapt to the current driving context (Sánchez et al., 2024).

6. Traffic-flow support and cooperative congestion control

SD is not restricted to the ego vehicle. In freeway traffic, support vehicles can be used to suppress congestion side effects created by jam-absorption driving (JAD). The cited study introduces support driving via connected and automated support vehicles (SVs) placed upstream of the absorbing vehicle. The SVs dynamically increase their time gaps according to a string-stability-based control law, with design guided by the head-to-tail string-stability criterion

$e(t)$04

In the reported simulations, the traffic stream contains $e(t)$05 vehicles, including one absorbing vehicle and five SVs, typically assigned every $e(t)$06 vehicles upstream (Suzuki et al., 5 Aug 2025).

The key result is that SD suppresses secondary shock waves caused by the absorbing vehicle’s deceleration, but does not itself eliminate the target shock wave. Relative to JAD-only, the combination of JAD and SD reduces total fuel consumption by $e(t)$07 kg and collision risk, measured by inverse time-to-collision, by $e(t)$08, while increasing total travel time by $e(t)$09 hours. A reversion strategy that returns the extended time gap to its initial value reduces travel time by $e(t)$10 hours relative to the non-reverting variant while maintaining low collision risk, albeit with $e(t)$11 kg higher fuel consumption. The study emphasizes that these gains are obtained with only six CAVs out of $e(t)$12 vehicles, consistent with low CAV penetration rates anticipated in early deployment stages (Suzuki et al., 5 Aug 2025).

This systems-level use of SD differs from haptic assistance or teleoperation, but it preserves the same structural idea: a support layer modifies local control authority or spacing behavior so that a larger dynamical system remains safe and efficient. The cited results therefore place SD within the broader literature on string stability, mixed-traffic control, and low-penetration CAV deployment (Suzuki et al., 5 Aug 2025).

7. Benchmarks, digital twins, and research directions

Recent SD research is strongly benchmark-driven. DrivIng provides a large-scale multimodal dataset with a complete geo-referenced digital twin of an approximately $e(t)$13 km route spanning urban, suburban, and highway segments. It includes six RGB cameras, one LiDAR, and ADMA-based localization; annotations at $e(t)$14 Hz comprise around $e(t)$15 million 3D instances across $e(t)$16 classes. The digital twin supports one-to-one transfer of real traffic into simulation through kinematic replay and interactive re-simulation. Benchmarking on nuScenes-style metrics shows PETR at day/dusk/night NDS of $e(t)$17 and CenterPoint at $e(t)$18, with both models degrading at night and the LiDAR-based model remaining substantially stronger (Rößle et al., 21 Jan 2026).

Hybrid-navigation benchmarking now explicitly measures SD–HD association. OMA is described as the first benchmark for associating global SD maps with online local HD maps for hybrid navigation. Built from nuScenes and OpenStreetMap, it contains more than $e(t)$19 scenarios, $e(t)$20 road segments, and $e(t)$21 lane paths. Its Association Precision-Recall metrics evaluate whether SD road-label sequences align with HD lane-path predictions, and the baseline Map Association Transformer uses path-aware attention, spatial attention, and topology-constrained beam search. The reported results show up to $e(t)$22 A-F1 gain on OMA and $e(t)$23 Association Precision gain on OMA-GT over classical baselines, with latency below $e(t)$24 ms per scene (Wan et al., 10 Jul 2025).

Open releases are a recurring theme. OpenLane-V2-OSM extends OpenLane-V2 with aligned OpenStreetMap data and supports both rasterized and graph SD-map representations, while SD-RouteFusion releases an SD-route generation toolkit for any dataset containing ego pose and future trajectories (Zhang et al., 2024, Voloshyn et al., 1 Jul 2026). This suggests that a large part of current SD progress is methodological standardization: common priors, common ODD descriptions, common safety metrics, and common replay or simulation infrastructures.

A broader synthesis of the literature indicates three durable research directions. First, support functions are moving from heuristic assistance toward explicit optimization under structured constraints, such as homography feasibility regions, trajectory-set Bayesian filtering, string-stability guarantees, or safety-centric route penalties (Zhong et al., 6 Jan 2026, Xia et al., 2024, Suzuki et al., 5 Aug 2025, Sahu et al., 29 May 2026). Second, scalable priors—SD maps, road manuals, route graphs, and digital twins—are being used to substitute for expensive HD-map infrastructure while preserving operational usefulness (Diwanji et al., 4 Feb 2025, Voloshyn et al., 1 Jul 2026, Rößle et al., 21 Jan 2026). Third, human factors remain central: learning curves, feedback modality, latency, and intervention structure continue to determine whether support is merely corrective or genuinely capability-expanding (Wada, 2019, Hans et al., 31 Mar 2025, Hans et al., 12 Mar 2025).