Closed-Loop Aware Planning via Imitation-Free Speed Scoring for End-to-End Autonomous Driving

1. Introduction

• Imitating a single expert demonstration can prevent end-to-end models from learning alternative valid behaviors in a multimodal driving scenario.The paper identifies this as a mismatch with closed-loop deployment.Source: §1 · p. 1
• Prior approaches either imitate one speed per class or combine driving criteria with imitation, while leaving test-time speed control unsupported.The evidence for criteria-based scoring benefits in reactive closed-loop settings is mixed, and path and speed often share an objective and architecture.Source: §1 · p. 1
• CLAP-AD imitates the path but replaces expert speed imitation with multi-criteria scoring over diverse speed classes.Candidate speeds along the predicted path provide safe speed profiles for deployment-time customization; criterion scores are supervised from physically grounded constraints rather than a single expert trace.Source: §1 · pp. 1–2
• CLAP-AD improves UniAD by +28.5 DS and VAD by +32.7 DS on Bench2Drive, and LTFv6 by +6.1 EPDMS on NAVSIM v2, while remaining within a few points of TFv6’s Bench2Drive baseline.Speed-Adherence improves on every backbone, by up to +25.2 points; only CLAP-AD variants reach adherence above 70.Source: §1 · p. 2, Figure 1 · p. 2

2.1. E2E Driving Architectures

• End-to-end architectures vary in scene representation and sensor-feature processing, but architectural gains on open-loop metrics have not consistently translated into strong Bench2Drive results.Examples include TransFuser’s camera–LiDAR fusion, UniAD’s query-based representation, and VAD’s vectorized representation.Source: §2.1 · p. 2

2.2. Planning Representation in E2E Models

• Open-loop imitation remains common because closed-loop training is computationally expensive, although single-trajectory demonstrations may not capture the multimodal behavior needed for interactive control.Few works target open-loop training strategies specifically for closed-loop performance.Source: §2.2 · p. 3
• Disentangled planning methods separate path and speed, but still imitate a single speed value or profile rather than model speed multimodality or offer test-time speed control.TF++, SimLingo, HiP-AD, and AlignDrive are examples of this line of work.Source: §2.2 · p. 3
• Trajectory-scoring methods add driving criteria alongside imitation, but reported reactive closed-loop benefits are mixed and some strongest results rely on imitation scoring alone.Hydra-NeXt’s added criteria yield marginal gains over imitation alone; SparseDriveV2’s strongest results rely on imitation scoring exclusively.Source: §2.2 · p. 3
• Prior user-customization methods rely on privileged agent states or training from scratch or per-style finetuning, unlike the paper’s stated goal of control in existing end-to-end models.The authors claim their approach is the first to provide test-time speed control without speed-conditioned training.Source: §2.2 · p. 3

3. Proposed Method

• CLAP-AD retains imitation learning for path prediction and uses scoring-based speed learning to represent driving’s speed-driven multimodality.The lightweight head integrates with existing end-to-end models for closed-loop-aware training in open-loop settings.Source: §3 · p. 3

3.1. Training Paradigm

• Candidate speed classes are scored for collision distance, whether red-light braking is achievable, and whether stop-sign braking is achievable, using targets derived from logged open-loop data.At inference, the network predicts scores from scene context without evaluating the underlying rule.Source: §3.1 · p. 3
• The collision target uses future logged agent positions and minimum distance along the predicted path; structured neighboring-path augmentation broadens collision supervision.Static objects are excluded because path prediction already avoids them.Source: §3.1 · p. 4
• Traffic-light and stop-sign targets reflect the deceleration required for a full stop, with scores decreasing as required braking exceeds comfortable levels and reaching zero when it is too high.When no stop line is ahead, all speeds are labeled safe.Source: §3.1 · p. 4

3.2. Planning Head Architecture

• The planning head predicts a path, then scores predefined speeds with three separate branches using the path and criterion-specific scene features; detached path inputs prevent scoring gradients from changing path learning.The collision branch uses dynamic-agent features, while traffic-light and stop-sign branches use their respective queries.Source: §3.2 · pp. 4–5, Figure 2 · p. 4

3.3. Controllability at Deployment

• At deployment, a speed is feasible when all criterion scores exceed a threshold, and a selector chooses among feasible classes by policy or proximity to the user’s requested speed.The request is re-evaluated each step; if no class is feasible, the selector falls back to 0 m/s. No retraining or architectural change is required.Source: §3.3 · p. 5

4.1. Benchmarks and Metrics

• Bench2Drive evaluates 220 closed-loop routes with DS, SR, and Efficiency; Bench2Drive-Speed adds 48 routes and requested-speed adherence, while NAVSIM v2 uses real-world pseudo-simulation and EPDMS.For baselines without speed requests, adherence measures how often their fixed speed coincides with the request, serving as a reference rather than controllability measure.Source: §4.1 · p. 5

4.2. Training Setup

• CLAP-AD replaces planning heads in UniAD, VAD, and TFv6; training spans Think2Drive, PDM-Lite, and human-driver demonstrations with differing consistency.Think2Drive is reinforcement-learned, PDM-Lite is rule-based, and NAVSIM experiments use human demonstrations.Source: §4.2 · pp. 5–6

4.3. Speed Adherence on Bench2Drive-Speed

• 82.53 SA gives CLAP-AD the best evaluated Speed-Adherence, 12.9 points above TFv6, for 4.1 DS below its baseline (88.47 vs 92.59).The comparison is on Bench2Drive-Speed; the TFv6 figures are its Driving Scores with and without CLAP-AD.Source: Table 1 · p. 6, §4.3 · p. 7

4.4. Closed-Loop Driving Performance on Bench2Drive

• On 220 Bench2Drive routes, CLAP-AD raises UniAD by +28.5 DS and VAD by +32.7 DS; its TFv6 result is 92.30 DS, 2.4 below baseline.Despite the TFv6 decrease, CLAP-AD surpasses the other Bench2Drive-trained methods compared in the passage.Source: Table 2 · p. 6, §4.4 · p. 7

4.5. Evaluation on NAVSIM v2

• On NAVSIM v2, CLAP-AD raises LTFv6 EPDMS by +6.1 (31.9 to 38.0), while larger-perception-backbone methods report higher EPDMS.Ego progress and extended comfort decline; the authors position CLAP-AD as competitive within its capacity regime, not state of the art.Source: Table 3 · p. 7, §4.5 · pp. 7–8

4.6. Ablation: Where the Gains Come From

• 59.44 DS versus 40.50 shows the same scoring objective benefits from CLAP-AD’s criterion-specific query head; path perturbations and longer horizons add +13 DS, with per-head encoders adding +2.5 DS.Extending horizons further to 24m and 2s does not help, which the authors attribute to noisier supervision that is less discriminative between neighboring speed classes.Source: Table 4 · p. 8, §4.6 · p. 8

5. Conclusion

• CLAP-AD uses an imitation-free, query-based multi-criteria scoring head for speed while training path and speed under different objectives.The approach builds on disentangled path and speed representations.Source: §5 · p. 8
• CLAP-AD improves UniAD and VAD by +28.5 and +32.7 Driving Score, stays within a few points of TFv6 on Bench2Drive, and surpasses its LTFv6 backbone by +6.1 EPDMS on NAVSIM v2.NAVSIM v2 is built on real-world driving data.Source: §5 · p. 8
• Explicit per-criterion safety scores support controller design and make decisions easier to audit, improving the practical deployability of end-to-end driving systems.The paper reports these advantages in its ablations.Source: §5 · p. 8

Supplementary Material

• The supplement covers planning-head architecture, test-time behavior control, per-ability results, collision ground truth, evaluation protocols, and limitations.It also reports NAVSIM v2 sub-metrics and trajectory conversion.Source: Supplementary Material · p. 12

A.1. Planning Head

• The planning head predicts a target-conditioned path and scores each speed class for collision risk and traffic-rule suitability using separate branches.The collision branch processes agents and the predicted path; traffic-light and stop-sign branches evaluate speed suitability.Source: §A.1 · p. 12
• On TFv6, CLAP-AD has 1.96M parameters and 0.137 GFLOPs per frame, about 3% of the model's 61.1M parameters.The speed classes are {0, 2.5, 5, 7, 8.5, 10, 12} m/s; collision horizons are 1 s for UniAD/VAD and 2 s for TFv6.Source: §A.1 · pp. 12–13

A.2. Backbone Query Selection

• The model filters detection and motion queries by predicted relevance, retaining dynamic agents and traffic elements for the corresponding scoring branches.TFv6 supplies a dense BEV map pooled into the fixed tokens expected by CLAP-AD's branches.Source: §A.1 · p. 12, §A.2 · p. 13

A.3. Backbone Auxiliary Task

• UniAD's perception detects traffic lights but lacks color predictions, so an auxiliary task predicts traffic-light color and whether lights or stop signs apply to the ego vehicle.The task addresses complex intersections with multiple visible controls; VAD and TFv6 require no such adaptation.Source: §A.3 · p. 13

A.4. Training Setup

• The supplementary classification head is trained on a curated 62,480-sample subset, then frozen for end-to-end training and inference.UniAD and VAD are fine-tuned end-to-end for one epoch; TFv6 is trained from scratch in two stages.Source: §A.3 · p. 13, §A.4 · p. 13

B. Test-Time Behavior Control on Bench2Drive

• Changing only the safe-score threshold and speed-selection rule spans roughly 77 Efficiency points without retraining; threshold 0.3 with median selection gives both backbones their highest Driving Score.Always selecting the fastest safe speed raises Efficiency but reduces UniAD's Driving Score from 74.29 to 69.53.Source: B. Test-Time Behavior Control on… · p. 13

C. Multi-Ability Results on Bench2Drive

• On UniAD and VAD, every ability improves; UniAD gains 31.1 points in overtaking and 46.7 in emergency braking, while traffic-sign scores rise from 14.21 to 55.26 on UniAD and 19.15 to 67.89 on VAD.The largest changes align with the abilities evaluated by the respective scoring heads.Source: C. Multi-Ability Results on Bench2Drive · p. 13, Table 7 · p. 14
• Traffic signs improve from 89.47 to 96.84 and give way from 40.00 to 50.00 on TFv6, while overtaking and emergency braking fall from 97.77 to 77.78 and 91.66 to 75.00.The paper attributes the declines to an already strong rule-based-expert baseline.Source: C. Multi-Ability Results on Bench2Drive · p. 14, Table 7 · p. 14

D. Collision Score Ground-Truth Calculation

• For each speed class, the collision target uses the minimum swept bounding-box distance between projected ego motion and actors' logged motion over the selected horizon.Distances above 1 m score fully safe, below 0.1 m score collision, and intermediate values are linearly interpolated.Source: D. Collision Score Ground-Truth… · p. 14

• Each criterion has its own scoring branch, so the scores reveal which constraint rejects a speed instead of leaving the decision implicit in a regressed trajectory.Source: E. Qualitative Results in CARLA · p. 14

F. NAVSIM Evaluation and Trajectory Generation

• NAVSIM v2 evaluates real logged data, then probes recovery using synthetic observations rendered from perturbed ego states rather than an interactive simulator.Source: F. NAVSIM Evaluation and Trajectory… · p. 14
• Stage 1 plans a four-second trajectory from the real observation; Stage 2 replans from roughly twelve synthetic observations per scene near the Stage 1 endpoint.Stage 2 scores are Gaussian-weighted toward viewpoints near the trajectory the agent predicted.Source: F. NAVSIM Evaluation and Trajectory… · pp. 14–15
• In the green-light scene with no stop sign or collision risk, all three scoring heads rate the full speed range highly, allowing any reasonable speed.Source: F. NAVSIM Evaluation and Trajectory… · p. 15, Figure 4 · p. 15
• EPDMS multiplies the two stage scores, requiring a method to plan well from the logged state and remain sensible from states it did not choose.Source: F. NAVSIM Evaluation and Trajectory… · p. 15
• Each stage combines nine sub-metrics: four rule-compliance measures enter multiplicatively, while five progress, safety, lane-keeping, and comfort measures enter a weighted average.A failure on any multiplicative measure zeroes that scene's score.Source: F. NAVSIM Evaluation and Trajectory… · p. 15
• CLAP-AD turns per-bin collision-safety logits into a continuous target speed by weighting speed bins with sigmoid confidence scores centered on a safety threshold.The soft weighting avoids abrupt target-speed changes under small changes in initial state, which Stage 2 viewpoints vary and extended comfort measures.Source: F. NAVSIM Evaluation and Trajectory… · p. 15
• The planner ramps from current speed toward the target under bounded acceleration and braking, then uses trapezoidal arc-length integration to resample the path into waypoints.If the ramp overshoots the distance budget, it binary-searches for the largest feasible ramp target; waypoint headings follow the local path tangent.Source: F. NAVSIM Evaluation and Trajectory… · pp. 15–16

G. Bench2Drive Evaluation Metrics

• Driving Score combines route completion with severity-weighted infraction penalties, while Success Rate measures routes completed without traffic violations within the allotted time.Source: G. Bench2Drive Evaluation Metrics · p. 16
• Efficiency measures speed relative to nearby traffic at 20 evenly spaced checkpoints, and Bench2Drive reports it independently from Driving Score.The CARLA Leaderboard instead folds its speed check into Driving Score as a penalty.Source: G. Bench2Drive Evaluation Metrics · p. 16
• Comfort is also reported by Bench2Drive, but its definition and discussion are handled separately.Source: G. Bench2Drive Evaluation Metrics · p. 16

H. On the Comfort Metric

• Bench2Drive Comfort can reward stationary or barely moving policies because it scores motion amplitudes, and the authors report that the metric tracks time spent stationary across models.Across eight models, stationary time and Comfort correlate at r = 0.97; a distance-based correction reduces the correlation to r = 0.06.Source: H. On the Comfort Metric · p. 16
• The highest Comfort score, 51.51, belongs to TCP-ctrl, one of the weakest drivers at 30.47 Driving Score, while stronger drivers cluster near 20.Across twenty methods, Comfort and Driving Score show a downward trend (r = −0.73).Source: H. On the Comfort Metric · p. 16
• Comfort thresholds fitted to ordinary human driving in nuPlan can exclude hard braking or sharp avoidance needed in Bench2Drive edge cases.Its hard Boolean conjunction also treats a marginal overshoot like a violent one, so it cannot distinguish controlled braking from erratic braking.Source: H. On the Comfort Metric · p. 17
• The authors therefore omit Comfort from planner comparisons and report Driving Score, Success Rate, and Efficiency instead.Source: H. On the Comfort Metric · p. 17

I. Limitations

• CLAP-AD scores collision, traffic lights, and stop signs, but not road speed limits; highway use would require a wider speed range and a speed-limit criterion.Scoring works best for hard rules with explicit, checkable targets; more complicated maneuvers may be difficult to formulate this way.Source: §I · p. 17