Open-loop (OL) to closed-loop (CL) gap (OL-CL gap) exists when OL-pretrained policies scoring high in OL evaluations fail to transfer effectively in closed-loop (CL) deployment. In this paper, we unveil the root causes of this systemic failure and propose a practical remedy. Specifically, we demonstrate that OL policies suffer from Observational Domain Shift and Objective Mismatch. We show that while the former is largely recoverable with adaptation techniques, the latter creates a structural inability to model complex reactive behaviors, which forms the primary OL-CL gap. We find that a wide range of OL policies learn a biased Q-value estimator that neglects both the reactive nature of CL simulations and the temporal awareness needed to reduce compounding errors. To this end, we propose a Test-Time Adaptation (TTA) framework that calibrates observational shift, reduces state-action biases, and enforces temporal consistency. Extensive experiments show that TTA effectively mitigates planning biases and yields superior scaling dynamics than its baseline counterparts. Furthermore, our analysis highlights the existence of blind spots in standard OL evaluation protocols that fail to capture the realities of closed-loop deployment.
Figures & tables
Figure 1: Motivation. (a) : Open-loop (OL) train-test paradigms do not necessarily align with the objectives and evaluation metrics in closed-loop (CL) deployment; (b) : This mismatch results in significant performance discrepancies, characterized by decaying correlation and diverging scaling dynamics; (c) : We attribute this OL-CL gap to observational domain shift and objective mismatch .
Name
Closed-loop
Long-term Simulation
MV-Consistency
TF-Consistency
Supported Maps
Supported Traffic Modes
NAVSIM [ 3 , 2 ]
×
×
✓
✓
nuPlan
Log-replay
Bench2Drive [ 6 ]
✓
✓
✓
✓
CARLA
IDM
DriveArena [ 27 ]
✓
✓
×
×
nuScenes
Log-replay
HUGSIM [ 28 ]
✓
✓
✓
×
nuScenes, Waymo, KITTI-360, PandaSet
Log-replay, IDM, Adversarial
BridgeSim (ours)
✓
✓
✓
✓
CARLA, Waymo, nuScenes, nuPlan, PG
Log-replay, IDM, Adversarial
Table 1: Comparisons with other E2E open-loop and closed-loop benchmarks. MV-Consistency: multi-view temporal consistency, TF-Consistency: Traffic elements annotations consistency. PG: procedural map and behavior generation. See Appendix B for more details of BridgeSim platform.
Figure 2: Empirical Study on Observational Domain Shift. (a) : Flow-matching procedure for observational calibration; (b) : Calibration effects on OL and CL metrics. See Appendix D.2 for experimental details.
Figure 3: Empirical Study on Objective Mismatch. (a) : Compounding execution error of J(π) w.r.t. execution horizon k ; (b) : J(π) under CL and OL objectives, where the divergence between these curves constitutes the objective mismatch gap ( Δobj ); (c) : OL-CL Pearson correlation under different traffic modes w.r.t. simulation horizon T . See Appendix D.3 for experimental details.
Figure 4: Test-time Policy Adaptation. (a) : Truncated Q-value estimator selects safe planned trajectories during closed-loop execution; (b) : Adaptive replan preserves previously verified planned trajectories if the newly proposed trajectory doesn’t obtain higher Q-values; (c) : Test-time scaling dynamics between biased (baseline) and truncated Q-value estimators (TTA).
Figure 5: Improvements of proposed TTA framework. (a): Performance gain under different simulation horizons on BridgeSim-NavHard scenarios; (b): Performance gain under different log-replay domains; (c): Component analysis of TTA framework.
Model
IDM
Safety-Critical
DS
EPDMS
RC
NC
DAC
TTC
LK
DS
EPDMS
RC
NC
DAC
TTC
LK
DiffusionDrive
45.30
72.95
59.93
95.41
85.47
89.73
89.83
38.68
68.16
54.77
93.62
83.86
84.97
90.71
DiffusionDriveV2
31.94
65.43
45.85
94.57
82.09
89.10
85.42
28.17
61.45
43.50
93.15
80.78
85.72
85.81
RAP
60.02
76.85
77.20
97.68
88.84
93.07
78.16
51.83
70.88
71.22
95.03
88.16
86.37
78.26
TTA (DiffusionDrive)
62.68 +17.4
84.36
72.28
96.85
95.63
91.57
92.43
55.02 +16.3
78.95
67.92
94.73
95.21
86.64
91.30
TTA (DiffusionDriveV2)
53.96 +22.0
77.47
67.89
96.08
90.53
89.46
89.98
47.64 +19.5
71.66
65.91
93.54
89.16
84.23
90.11
Table 2: Performance comparisons with different reactive traffic modes. Note that TTA improves key closed-loop metrics such as NC , TTC , and RC , resulting in overall higher DS . See Appendix D.4 for complete sub-metric scores.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Illustration of BridgeSim Platform. BridgeSim supports unified closed-loop simulation of policies pretrained from any other simulators, enabling us to perform rigorous decomposition analysis.
Metric
Description
Critical Constraints ( C )
No At-Fault Collisions ( NC )
Ego polygon Pego must not intersect any replayed object polygons Pobj .
Drivable Area Compliance ( DAC )
Entire horizon must satisfy Pego⊂Mdrivable within the HD map.
Traffic Light Compliance ( TLC )
Agent cannot cross stop lines associated with red light states.
Driving Direction Compliance ( DDC )
Heading deviation ∣θego−θlane∣ must remain within ±π/2 .
Quality Features ( F )
Appendix
Table 3: Descriptions of subscores of EPDMS .
Model
Pretrained Sim
DS
EPDMS
RC
NC
DAC
DDC
TLC
TTC
LK
HC
EC
Human Expert
N/A
91.14
91.96
98.86
98.18
99.23
99.76
98.62
94.55
90.22
99.10
71.44
UniAD (CVPR’23)
Bench2Drive
31.00
57.55
55.25
93.15
71.75
99.76
99.61
81.13
97.70
74.25
30.04
VAD (ICCV’23)
Bench2Drive
26.25
57.60
46.76
93.29
71.22
99.76
99.68
82.31
98.67
71.13
27.67
DiffusionDrive (CVPR’25)
NAVSIM
43.29
69.70
61.30
95.38
81.48
99.76
98.81
87.47
91.57
91.64
39.02
DiffusionDriveV2 (arXiv)
NAVSIM
29.81
62.01
45.85
93.48
79.75
99.76
98.83
86.93
84.11
71.38
20.62
RAP (ICLR’26)
NAVSIM
52.08
72.41
70.74
95.63
85.24
99.76
99.18
88.14
84.69
92.84
48.91
Appendix
Table 4: Quantitative comparison of E2E driving policies evaluated on BridgeSim-NavHard. All results are reported with a replan rate of 5 and a simulation horizon of 8 seconds. We use rr=5 to align with the temporal discretization commonly used for training on NAVSIM/OpenScene, where the dataset is provided at 2 Hz (i.e., Δt=0.5s ).
Figure 7: Visualization of observation pairs in source and target domains.
Figure 8: Qualitative analysis of simulation artifacts in DriveArena: multi-view incoherence (left) and temporal instability (right). Red boxes highlight inconsistencies that lead to policy failures or metric distortion.
Figure 9: Qualitative analysis of reconstruction artifacts and ghosting effects in HUGSIM. Red boxes highlight inconsistencies.
Model
IDM
Safety-Critical
DS
EPDMS
RC
NC
DAC
DDC
TL
TTC
LK
HC
EC
DS
EPDM
RC
NC
DAC
DDC
TL
TTC
LK
HC
EC
DiffusionDrive
45.30
72.95
59.93
95.41
85.47
100.00
98.36
89.73
89.83
92.31
40.82
38.68
68.16
54.77
93.62
83.86
100.00
98.36
84.97
90.71
89.56
39.13
DiffusionDriveV2
31.94
65.43
45.85
94.57
82.09
100.00
98.23
89.10
85.42
74.52
22.99
28.17
61.45
43.50
93.15
80.78
100.00
98.30
85.72
85.81
70.99
21.68
RAP
60.02
76.85
77.20
97.68
88.84
100.00
98.49
93.07
78.16
94.49
53.84
51.83
70.88
71.22
95.03
88.16
100.00
98.58
86.37
78.26
90.22
50.53
TTA (DiffusionDrive)
62.68 +17.4
84.36
72.28
96.85
95.63
100.00
98.98
91.57
92.43
87.95
40.13
55.02 +16.3
78.95
67.92
94.73
95.21
100.00
98.76
86.64
91.30
83.65
37.11
TTA (DiffusionDriveV2)
53.96 +22.0
77.47
67.89
96.08
90.53
100.00
99.43
89.46
89.98
85.46
29.90
47.64 +19.5
71.66
65.91
93.54
89.16
100.00
99.07
84.23
90.11
81.73
27.49
Appendix
Table 5: Performance comparisons with different reactive traffic modes with complete sub-metric scores.
Open-loop imitation learning has advanced modern autonomous driving policy architectures, but closed-loop deployment remains vulnerable to policy-induced distribution shift. Existing post-training paradigms exhibit fundamental trade-offs: closed-loop RL fine-tuning provides grounded feedback from executed actions but is constrained by the sparsity of informative events, whereas counterfactual fine-tuning provides dense supervision over candidate futures but inherits bias from imperfect future estimates. We introduce Counterfactual-to-Interactive Reinforcement Fine-Tuning (CRAFT), an on-policy framework that formulates closed-loop post-training as proxy-residual optimization. CRAFT uses group-normalized counterfactual advantages as a dense proxy for real closed-loop advantages and aligns this proxy with the closed-loop world through grounded residual correction from interaction-critical events. To stabilize adaptation, CRAFT regularizes the online policy toward an EMA teacher via asymmetric KL self-distillation. Theoretically, CRAFT decomposes the real closed-loop policy gradient into proxy and residual terms under the same visited-state distribution, reducing residual variance with an aligned proxy while mitigating proxy bias through grounded residual approximation. Empirically, CRAFT achieves the strongest closed-loop gains on Bench2Drive across hierarchical planning, vision-language-action, and vocabulary-scoring architectures. Ablations, scaling behavior, stability analyses, and transfer results further validate the complementary roles of dense counterfactual proxy and grounded residual correction. Project page: https://currychen77.github.io/CRAFT.
Keyu Chen, Nanfei Ye, Yida Wang +4
School of Vehicle and Mobility, Tsinghua University · Li Auto Inc
End-to-end (E2E) driving policies have advanced rapidly on open-loop (OL) benchmarks, yet OL evaluation cannot reveal whether a policy can withstand compounding errors, recover from failures, or interact safely with surrounding actors. We introduce NavSafe-∞, a photorealistic closed-loop (CL) benchmark comprising 280 scenarios spanning 28 event types, each with success and failure criteria defined within a structured traffic-safety taxonomy, yielding category-level capability scores for Traffic Crashes, Vulnerable Road User Crashes, Traffic Violations, and Traffic Incidents. After evaluating 20 E2E policies, we find that OL gains do not reliably transfer to CL safety. Analysis of two common remedies reveals that (1) passive demonstration perturbation helps mainly when CL rollouts stay near their perturbed training states, and (2) OL reinforcement-learning fine-tuning exhibits reward hacking by trading safety margin for ego progress, which CL feedback amplifies into compounding safety-critical errors. Together, these results demonstrate the blind spot of OL benchmarks indicating CL safety success. The benchmark and an extensible toolbox for customizable event curation and policy diagnosis will be open-sourced to facilitate future research.
Generalization under distribution shift remains a central bottleneck for closed-loop autonomous driving. Although simulators like CARLA enable safe and scalable testing, existing benchmarks rarely measure true generalization: they typically reuse training scenarios at test time. Success can therefore reflect memorization rather than robust driving behavior. We introduce Fail2Drive, the first paired-route benchmark for closed-loop generalization in CARLA, with 200 routes and 17 new scenario classes spanning appearance, layout, behavioral, and robustness shifts. Each shifted route is matched with an in-distribution counterpart, isolating the effect of the shift and turning qualitative failures into quantitative diagnostics. Evaluating multiple state-of-the-art models reveals consistent degradation, with an average success-rate drop of 22.8%. Our analysis uncovers unexpected failure modes, such as ignoring objects clearly visible in the LiDAR and failing to learn the fundamental concepts of free and occupied space. To accelerate follow-up work, Fail2Drive includes an open-source toolbox for creating new scenarios and validating solvability via a privileged expert policy. Together, these components establish a reproducible foundation for benchmarking and improving closed-loop driving generalization. We open-source all code, data, and tools at https://github.com/autonomousvision/fail2drive .