Once deployed on vehicles, end-to-end driving models continuously encounter challenging scenarios, generating valuable human takeover data that capture both model outputs and human behaviors. However, human interventions are often imperfect, as drivers may intervene too late or unnecessarily. Consequently, direct imitation can reinforce late responses or discourage otherwise appropriate maneuvers. To address this challenge, we propose an offline post-training framework that transforms imperfect takeover data into proactive supervision. We first construct diverse trajectory candidates by augmenting the plans of the deployed model and shifting recorded human corrective maneuvers earlier in time. We then train a scene-conditioned trajectory scorer to rank these candidates using expert demonstrations and unpaired preference signals from takeover logs. Using this scorer, we selectively replace pre-takeover training targets with higher-scoring candidates while retaining the original targets in normal driving scenarios. Finally, we distill the resulting targets into the driving model without changing its inference architecture. In closed-loop evaluations using real-world scenarios reconstructed with 4D Gaussian splatting, our method achieves the highest overall success rate, outperforming the deployed model, direct imitation learning and trajectory post-optimization. These results highlight a practical path toward efficient updates of deployed driving models using existing fleet logs.
Figures & tables
Figure 1: Overview of the proposed offline post-training framework. Left : Real-world takeover logs capture the model’s planning results and subsequent human corrections. Middle : Diverse trajectory candidates are constructed through temporal advancement of human corrections and random augmentation of model trajectories. A scene-conditioned trajectory scorer is trained using expert trajectory matching and unpaired preference supervision. Right : The learned scorer selectively refines pre-takeover supervision targets, which are distilled into the E2E driving model without modifying its inference architecture.
Figure 2: Results of trajectory augmentation in four representative real-world takeover scenarios. Each case shows the onboard camera image (top) and augmented trajectories in BEV (bottom). (a) Failure to follow right-turn navigation. (b) Insufficient avoidance of a sanitation worker. (c) Potential collision during a nighttime right turn. (d) Curb collision risk on a sharply curved ramp.
Figure 3: Qualitative comparison of supervision targets in the four takeover scenarios shown in Figure 2 . Red, green, and orange denote the recorded ground truth, our scorer-refined target, and the post-optimization target, respectively.
Method
Hit@1
ADE@3s
ADE@6s
FDE@3s
FDE@6s
Oracle †
100%
0.156
0.327
0.299
0.832
Random
0.86%
1.879
5.001
4.231
12.221
Rule-based
0.32%
1.432
3.758
3.229
8.914
PP-only
12.65%
0.486
1.734
1.343
4.722
CE-only
9.37%
0.488
1.745
1.360
4.765
CE+PP
8.04%
0.560
1.911
1.592
5.213
Table 1: Expert-aligned trajectory selection on the expert driving set. All methods rank the same 133 candidates.
Method
Cut-in
Meeting
Bypass
Junction
ULT
Overall †
Base
86.94
61.40
40.82
67.72
42.55
70.87
Direct IL
84.90
78.95
38.78
74.07
63.83
75.30
Self-distillation
84.90
73.68
48.98
68.78
57.45
73.42
Post-optimization
81.63
84.21
40.82
70.90
42.55
71.89
Rule Refinement
86.12
75.44
48.98
71.43
59.57
75.13
Reward Finetune
88.16
73.68
36.73
76.72
70.21
77.34
Table 2: Closed-loop success rate across 5 challenging scenario categories. Bypass denotes obstacle avoidance maneuvers constrained by oncoming traffic.
Figure 4: Qualitative comparison in a 4DGS-based closed-loop pedestrian avoidance scenario. The top and bottom groups show the deployed base model and our post-trained model, respectively. The left panels show two key camera frames ( T1 and T2 ) selected separately for each model, with the corresponding acceleration and vehicle speed on the right.
Setting
Naug
Ne
Overall(%)
Random augmentation
0
5
73.23
5
5
78.62
Correction shifting
3
1
75.25
3
3
75.93
Full (Ours)
3
5
81.48
Table 3: Effect of trajectory augmentation on the overall closed-loop success rate. All success rates are calculated over 594 valid scenarios.
Setting
CE+PP
CE-only
Ours+UR
Ours
Success rate (%)
76.31
77.16
79.36
81.39
GT replacement rate (%)
28.98
39.81
77.87
34.39
Table 4: Effect of scorer supervision on target refinement and closed-loop success rate. All settings are evaluated on the same 591 valid cases.
Autonomous driving develops vehicles that perceive road scenes, predict interactions, and plan and control safe movement with limited or no human input.
Current Vision-Language-Action (VLA) paradigms in end-to-end autonomous driving rely on offline training from static datasets, leaving them vulnerable to distribution shift. Recent post-training methods use takeover data to mitigate this by augmenting the dataset with high-quality expert takeover samples, yet they suffer from two key limitations: supervision restricted to the period after the takeover moments leads to policies with limited safety margins, and passive preference optimization lacks active exploration for optimal performance. In this paper, we propose TakeVLA, a novel VLA post-training framework that overcomes these shortcomings through two complementary innovations. First, we introduce pre-takeover language supervision, which allows the VLA to learn from mistakes proactively. By explicitly teaching the model about what to do in error-prone situations, we cultivate a precautionary mindset that anticipates hazards early and substantially enlarges safety margins. Second, we propose Scenario Dreaming, a reinforcement fine-tuning paradigm that operates in reconstruceted takeover scenarios, encouraging active exploration beyond mere preference fitting. Experiments on the Bench2Drive benchmark demonstrate that TakeVLA achieves state-of-the-art closed-loop performance, surpassing the strong VLA baseline SimLingo by 4.93 in driving score, with an enhanced safety margin as evidenced by an 11.76% increase in average TTC.
Yinfeng Gao, Deqing Liu, Qichao Zhang +7
University of Science and Technology Beijing · Xiaomi EV · Institute of Automation, Chinese Academy of Sciences
End-to-end models that map multimodal inputs directly to future trajectories/maneuvers have emerged as an increasingly prominent research paradigm in autonomous driving. This class of models includes both Vision-Language-Action models and trajectory-generative planners. Unlike classic machine learning applications, autonomous vehicles operate in safety-critical and interaction-intensive environments where traditional open-loop imitation of expert demonstrations is not sufficient to ensure reliability. In particular, small execution errors can accumulate over time, while recovery behaviors are scarce in training data. In addition, long-horizon objectives such as safety and driving comfort are not captured by pointwise labels either. These limitations have motivated a shift toward post-training techniques, which further refine driving policies beyond pure imitation. This survey presents a unified view of post-training for autonomous driving by defining its scope and organizing the existing literature into four major families based on the form of supervision they use. For each family, we discuss its capabilities, limitations, and open challenges. We aim to facilitate a systematic understanding of this emerging area and stimulate future research on reliable and efficient post-training for autonomous driving.A collection of related papers is available at https://github.com/RYNing/Awesome-Post-Training-In-Autonomous-Driving-Papers.
Ruining Yang, Muxing Wang, Yixiao Chen +8
Northeastern University, Boston, USA · Purdue University, West Lafayette, USA
As scaling pre-training data alone yields diminishing returns, post-training is becoming increasingly important across physical AI domains such as autonomous driving. End-to-end driving policies are pre-trained in open loop with behavior cloning on human demonstrations. However, compounding errors during closed-loop deployment can take the vehicle outside the training data distribution, increasing the risk of safety-critical incidents. Closed-loop post-training can mitigate this risk but requires costly simulation for sensor-based policies. We propose OPTED (on-policy fine-tuning for end-to-end driving) which decouples reinforcement learning from the post-training of the end-to-end policy: a privileged teacher is trained using RL on vectorized inputs (HD-map and bounding boxes). This teacher then provides supervision to the pre-trained student during closed-loop post-training. We apply OPTED to two camera-based models, TransFuser and VaVAM, and fine-tune them in AlpaSim, using neural reconstructions (3DGS) of real driving logs. Driving scores increase by factors of 1.6× and 9.5×, respectively. In controlled experiments OPTED matches closed-loop performance with approximately three orders of magnitude fewer simulator interactions than direct RL post-training, while staying closer to the human prior. Project page: https://01dami23.github.io/opted/