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

Emad Aghajanzadeh; Thomas Schwarzfischer; Marc Kaufeld; Moritz Thoma; Ryan Pégoud; Manoj Rohit Vemparala; Pierpaolo Mori; Nael Fasfous

cs.ROOct 26, 2026

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

Authors: Emad Aghajanzadeh, Thomas Schwarzfischer, Marc Kaufeld, Moritz Thoma, Ryan Pégoud, Manoj Rohit Vemparala, Pierpaolo Mori, Nael Fasfous,

Organizations: Technical University of Munich, Germany · BMW Group, Germany · TU Wien, Austria

Abstract

End-to-end autonomous driving has predominantly relied on open-loop imitation learning for training, which presents two limitations. First, expert demonstrations show only one behavior per scenario, so the model cannot learn alternative valid behaviors and, in turn, cannot adapt to user-specified preferences at test time. Second, noise or ambiguity in these demonstrations can produce inconsistent, unreliable supervision for the end-to-end models. Driving is inherently multi-modal: multiple different behaviors can be valid responses to the same scene, and this diversity arises largely from speed rather than path. We introduce a planning paradigm that learns the path through imitation while scoring a set of candidate speeds across multiple criteria with a learned function supervised by physically grounded safety constraints. We realize this paradigm as Closed-Loop Aware Planning for Autonomous Driving (CLAP-AD), a plug-in planning head that integrates with existing end-to-end driving models and, to our knowledge, is the first to enable explicit test-time control over their speed selection without speed-conditioned training. On Bench2Drive-Speed, CLAP-AD sets a new state of the art Speed-Adherence Score of 82.53, topping TFv6, the strongest baseline on this metric, by 12.9 points while retaining strong closed-loop driving performance. On Bench2Drive, it improves driving score by +28.5 and +32.7 over UniAD and VAD, and on NAVSIM v2, built on real-world driving data, EPDMS by +6.1 over its backbone. Project page: https://clap-ad.github.io

Figures & tables

Panel of Figure 1 — Figure 1: CLAP-AD improves speed-adherence across three different end-to-end driving models. 1(a) CLAP-AD consistently improves Speed-Adherence across all three backbones; only on TFv6 does this cost a small amount of Driving Score. 1(b) An example of how CLAP-AD respects the user’s requested target speed.

Figure 2: Overview of the CLAP-AD planning head. The architecture consists of an Ego-Path block that takes ego detection and motion queries as input and predicts a feasible driving path. The Collision block encodes surrounding agents’ detection and motion queries, fuses them with the predicted path representation, and uses learned collision queries to produce collision scores. Similarly, the Traffic Light and Stop Sign blocks employ dedicated queries that attend to their respective detection features together with the path embedding to generate their speed scores.

Panel of Figure 3 — Figure 3: Speed-profile variability across two experts on a repeated HardBrake scenario. Lines show the mean speed over three re-runs of the same route; shaded bands show $\pm 1$ standard deviation. PDM-Lite and Think2Drive brake at markedly different points along the route, showing that even expert-level speed behavior is inherently multi-modal. Within each expert, PDM-Lite (rule-based) reproduces an identical profile across all three runs (avg. std $=0.00$ m/s), while Think2Drive (RL-based) diverges substantially between runs (avg. std $=3.02$ m/s).

	Speed-Adherence $\uparrow$				Traditional Metrics $\uparrow$
Method	All	Easy	Medium	Hard	Efficiency	Driving Score	Success Rate
TCP-Speed [ 35 ]	69.23	76.18	68.71	62.81	169.84	29.66	18.75
Hydra-NeXt [ 27 ]	68.53	68.68	67.94	68.96	157.05	69.87	39.58
SparseDriveV2 [ 40 ]	51.13	52.50	49.72	51.18	91.77	94.63	60.41
UniAD [ 11 ]	54.22	56.14	52.62	53.90	108.49	18.90	4.16
+ CLAP-AD (ours)	79.41 $+$ 25.2	83.83	75.40	79.00	170.03	75.15 $+$ 56.3	41.67

Table 1: Bench2Drive-Speed results on 48 routes. Speed-Adherence (SA) measures how faithfully a policy follows the user-specified speed. Each backbone sits directly above its CLAP-AD variant; inline values give the change relative to it. Best per column in bold . Shaded cells mark Speed-Adherence above 70, the controllable range highlighted in Fig. 1(a) .

Bench2Drive-base data (Think2Drive [ 25 ] expert)
Method	Ctrl.	DS $\uparrow$	$\Delta$ DS	SR (%) $\uparrow$
SparseDrive $\dagger$ [ 39 ]	✗	44.54	–	16.71
GenAD [ 50 ]	✗	44.81	–	15.90
MomAD [ 37 ]	✗	47.91	–	18.11
DiFSD [ 38 ]	✗	52.02	–	21.00
Hydra-MDP [ 26 ]	✗	59.95	–	29.82

Table 2: Closed-loop results on the 220 Bench2Drive routes. Methods are grouped by training expert, each backbone directly above its CLAP-AD variant. Ctrl. : supports user-specified speed at test time.

\Delta

DS is relative to the method’s own backbone.

\dagger

from the re-implementation of [ 37 ] . Shaded rows are the only methods offering test-time speed control; Tab. 1 reports the Speed-Adherence gained in exchange for the small DS cost on TFv6.

Method	Stage	NC $\uparrow$	DAC $\uparrow$	DDC $\uparrow$	TL $\uparrow$	EP $\uparrow$	TTC $\uparrow$	LK $\uparrow$	HC $\uparrow$	EC $\uparrow$	EPDMS $\uparrow$
TransFuser [ 7 ]	Stage 1	96.2	79.5	99.1	99.5	84.1	95.1	94.2	97.5	79.1	23.1
TransFuser [ 7 ]	Stage 2	77.7	70.2	84.2	98.0	85.1	75.6	45.4	95.7	75.9	23.1
DiffusionDrive [ 28 ]	Stage 1	96.0	79.7	97.4	99.5	81.3	93.1	90.8	96.8	73.8	24.2
DiffusionDrive [ 28 ]	Stage 2	82.1	72.2	88.5	98.7	85.1	78.8	49.2	89.3	71.2	24.2
GuideFlow [ 29 ]	Stage 1	96.6	80.5	96.3	99.3	82.3	94.9	91.5	97.7	67.8	27.1
GuideFlow [ 29 ]	Stage 2	87.3	76.7	88.8	99.2	84.3	85.1	49.7	93.1	44.5	27.1

Table 3: Results on the NAVSIM v2 navhard benchmark [ 5 ] . CLAP-AD replaces only the planning head of LTFv6 and shares its backbone, so the shaded rows should be read directly against the LTFv6 rows above. As in Tab. 2 , shading marks the model offering test-time speed control. Results for LEAD-LTFv6 are taken from the public NAVSIM v2 leaderboard.

Speed objective	Planning head	DS $\uparrow$	SR (%) $\uparrow$
Joint temporal wp. (baseline)	UniAD	45.12	20
Path + speed classification [ 12 ]	UniAD	43.03	0
Dual spatial–temporal wp. [ 33 , 43 ]	UniAD	56.44	20
Multi-criteria scoring (ours)	UniAD	40.50	0
Multi-criteria scoring (ours)	CLAP-AD	59.44	30
+ path aug. & horizon (12 m / 1 s)	CLAP-AD	72.71	50

Table 4: Ablation on UniAD, evaluated on Dev10. The upper block varies only the speed objective inside UniAD’s original planning head; the lower block keeps our scoring objective and replaces the head with CLAP-AD. Scoring only pays off once the head is designed to consume it (40.50

\rightarrow

59.44).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Module	Architecture
Ego–Path Block
Plan Fuser	FC( $256\!\cdot\!fuser\_dim$ , 256)
Path Decoder	TP_FC(2,256), GRU(256,256) $\rightarrow$ FC(256,2)
Collision Branch
Path Encoder	MLP(256,512,256,24)
Agent Fuser	FC( $256\!\cdot\!fuser\_dim$ ,256)

Appendix

Table 5: Detailed architecture of CLAP-AD. The fusion dimension

fuser\_dim

is 3 for UniAD and 4 for VAD. FC(a,b) denotes Linear(a,b)

\rightarrow

\rightarrow

ReLU. MLP(a,b,c) denotes Linear(a,b)

\rightarrow

\rightarrow

ReLU

\rightarrow

Linear(b,c) , while MLP(a,b,c,d) denotes Linear(a,b)

\rightarrow

ReLU

\rightarrow

Linear(b,c)

\rightarrow

ReLU

\rightarrow

Linear(c,d) . Attention refers to the attention module used to aggregate contextual features.

Backbone	Score	Speed	Eff.	Driving	Success
	Thresh.	Sel.	$\uparrow$	Score $\uparrow$	Rate(%) $\uparrow$
UniAD	0.5	median	173.55	73.57	47.94
	0.3	median	176.03	74.29	49.09
	0.3	max	252.84	69.53	40.45
VAD	0.5	median	172.56	71.21	42.73
	0.3	median	179.01	75.04	47.27

Appendix

Table 6: Test-time behavior control on the 220 Bench2Drive routes. One trained model under different controller settings. The threshold sets which scores count as safe; the selection rule picks the target speed from the safe set. Varying only these inference-time parameters spans cautious to throughput-oriented driving, with no retraining.

Method	Merging	Overtaking	Emerg. Brake	Give Way	Traffic Sign	Mean
UniAD	14.10	17.78	21.67	10.00	14.21	15.55
CLAP-AD on UniAD	31.25	48.88	68.33	40.00	55.26	48.74
VAD	8.11	24.44	18.64	20.00	19.15	18.07
CLAP-AD on VAD	40.00	42.22	55.00	60.00	67.89	53.02
TFv6	72.50	97.77	91.66	40.00	89.47	78.28
CLAP-AD on TFv6	71.25	77.78	75.00	50.00	96.84	74.17

Appendix

Table 7: Multi-ability results on Bench2Drive (%). On the Bench2Drive-trained backbones CLAP-AD improves every ability. On TFv6, trained on a rule-based expert, the traffic-rule abilities still improve while the collision-driven ones trade against an already strong baseline.

Panel of Figure 4 — Figure 4: Front camera images of different driving scenarios with predicted speed scores. Collision, traffic light, and stop sign branches individually predict scores for every speed class.

Panel of Figure 5 — Figure 5: Driving Score versus Comfort on the 220 Bench2Drive routes. Comfort falls as Driving Score rises; the dashed line is a least-squares fit over the methods shown. This is not evidence that competent driving is uncomfortable: every quantity the metric monitors is an amplitude of motion, so policies that stall or creep satisfy all of them trivially.

Autonomous Driving Planning

Autonomous driving planning selects safe, feasible vehicle actions and trajectories amid traffic, uncertainty, road rules, and changing goals.

See all 384 papers

CLOVER: Closed-Loop Value Estimation and Ranking for End-to-End Autonomous Driving Planning
May 14, 2026Sining Ang, Yuguang Yang, Canyu Chen +1End-to-End Autonomous Driving Value Function Estimation
Department of Automation, University of Science and Technology of China · Institute for AI Industry Research, Tsinghua University · School of Electronic Information Engineering, Beihang University +1
What Matters for Scalable and Robust Learning in End-to-End Driving Planners?
Mar 16, 2026David Holtz, Niklas Hanselmann, Simon Doll +2End-to-End Autonomous Driving Autonomous Driving Planning
Mercedes-Benz AG · Max-Planck-Institute for Informatics, SIC
DriveMLP: When Training Strategy Outperforms Model Complexity in Autonomous Driving
Oct 21, 2026Alessandro Canevaro, Julian Schmidt, Hang Yu +4Autonomous Driving Planning Imitation Learning
Mercedes-Benz AG, Research & Development, Sindelfingen, Germany · University of Tübingen, Tübingen, Germany · Karlsruhe Institute of Technology, ITIV, Karlsruhe, Germany +2

May 14, 2026J/K move · Enter open · S save

CLOVER: Closed-Loop Value Estimation and Ranking for End-to-End Autonomous Driving Planning

Sining Ang, Yuguang Yang, Canyu Chen, Yan Wang

Department of Automation, University of Science and Technology of China · Institute for AI Industry Research, Tsinghua University · School of Electronic Information Engineering, Beihang University+1

End-to-End Autonomous Driving Value Function Estimation NavSim Autonomous Driving Planning

Open paper AskPDF arXiv

End-to-end autonomous driving planners are commonly trained by imitating a single logged trajectory, yet evaluated by rule-based planning metrics that measure safety, feasibility, progress, and comfort. This creates a training--evaluation mismatch: trajectories close to the logged path may violate planning rules, while alternatives farther from the demonstration can remain valid and high-scoring. The mismatch is especially limiting for proposal-selection planners, whose performance depends on candidate-set coverage and scorer ranking quality. We propose CLOVER, a Closed-LOop Value Estimation and Ranking framework for end-to-end autonomous driving planning. CLOVER follows a lightweight generator--scorer formulation: a generator produces diverse candidate trajectories, and a scorer predicts planning-metric sub-scores to rank them at inference time. To expand proposal support beyond single-trajectory imitation, CLOVER constructs evaluator-filtered pseudo-expert trajectories and trains the generator with set-level coverage supervision. It then performs conservative closed-loop self-distillation: the scorer is fitted to true evaluator sub-scores on generated proposals, while the generator is refined toward teacher-selected top-

k

and vector-Pareto targets with stability regularization. We analyze when an imperfect scorer can improve the generator, showing that scorer-mediated refinement is reliable when scorer-selected targets are enriched under the true evaluator and updates remain conservative. On NAVSIM, CLOVER achieves 94.5 PDMS and 90.4 EPDMS, establishing a new state of the art. On the more challenging NavHard split, it obtains 48.3 EPDMS, matching the strongest reported result. On supplementary nuScenes open-loop evaluation, CLOVER achieves the lowest L2 error and collision rate among compared methods. Code data will be released at https://github.com/WilliamXuanYu/CLOVER.

Abstract

Autonomous Driving Planning

Explore similar work

CLOVER: Closed-Loop Value Estimation and Ranking for End-to-End Autonomous Driving Planning

What Matters for Scalable and Robust Learning in End-to-End Driving Planners?

DriveMLP: When Training Strategy Outperforms Model Complexity in Autonomous Driving