ZTRS: Zero-Imitation End-to-end Driving

• ZTRS is an end-to-end autonomous driving framework that bypasses imitation learning by scoring a fixed set of candidate trajectories using offline reward signals.
• It integrates a deep sensor-driven image backbone, trajectory tokenization with Transformer Decoder, and employs Exhaustive Policy Optimization for dense supervision.
• ZTRS achieves state-of-the-art performance on benchmarks like Navtest, Navhard, and HUGSIM, balancing safety, progress, and comfort through reward-based optimization.

to=arxiv_search.search ￣奇米影视json {"query":"(Li et al., 28 Oct 2025) ZTRS Zero-Imitation End-to-end Autonomous Driving with Trajectory Scoring", "max_results": 5} to=arxiv_search.search .| json {"query":"HUGSIM autonomous driving benchmark arXiv", "max_results": 5} ZTRS, short for Zero-Imitation End-to-end Autonomous Driving with Trajectory Scoring, is an end-to-end autonomous driving framework that maps high-dimensional sensor inputs directly to ego-vehicle trajectories while removing imitation learning from the training loop entirely. Instead of regressing a continuous path or mimicking expert demonstrations, it scores a fixed, enumerable set of candidate trajectories using offline reward signals, combining end-to-end perception with offline reinforcement learning. The framework is introduced as the first method in this setting that learns only from rewards while operating directly on raw sensor data such as front-view images, and it is evaluated on open-loop and closed-loop driving benchmarks including Navtest, Navhard, and HUGSIM (Li et al., 28 Oct 2025).

1. Conceptual position within end-to-end driving

The motivation for ZTRS is the division of prior end-to-end driving systems into two main paradigms. Imitation learning methods are effective on real-world sensor data, but they inherit suboptimal expert behavior and are vulnerable to covariate shift when deployment states depart from the demonstrated distribution. Reinforcement learning methods can optimize directly for reward and scale in simulation, but prior driving systems of that type have usually relied on low-dimensional symbolic inputs such as object lists, traffic signals, or HD maps rather than raw sensor streams. ZTRS is designed to bridge that split by retaining sensor-rich end-to-end perception while replacing imitation targets with offline reward optimization (Li et al., 28 Oct 2025).

A defining modeling choice is to cast driving as trajectory scoring over a finite action set,

$\mathcal{A} = \{a_i\}_{i=1}^n,$

where each $a_i$ is a 4-second trajectory candidate. This replaces continuous trajectory regression with ranking over a fixed trajectory vocabulary. The paper argues that this makes reward evaluation feasible in the offline setting and avoids the cold-start difficulty associated with exploration in continuous action spaces.

ZTRS is therefore “zero-imitation” in a specific technical sense: the planner does not use human trajectory labels as supervision. The training corpus is an offline driving dataset, and the supervisory signal is derived from reward computations on the candidate trajectory set. This differentiates ZTRS from trajectory-scoring methods that still rely on imitation targets or human demonstrations.

2. Architecture and perception-to-trajectory pipeline

ZTRS is organized as a trajectory scorer with five modules: Image backbone, Trajectory tokenizer, Transformer Decoder, Policy head, and Multiple scoring heads (Li et al., 28 Oct 2025).

For a sampled state $s$ from the offline dataset $\mathcal{D}$, where $s$ contains sensor data and ego-vehicle status, the forward process is structured as follows. The image backbone extracts $L$ image tokens from the input image. The input image is a frontal-view image constructed by concatenating center-cropped front-left and front-right views. In parallel, the trajectory tokenizer encodes the fixed trajectory candidates into trajectory queries. These trajectory queries then attend to the image tokens inside a Transformer Decoder, so that each candidate trajectory token is contextualized by the visual scene.

The output stage separates planning likelihood from metric prediction. The policy head produces a distribution $\pi(\cdot \mid s)$ over the trajectory set, while the scoring heads produce metric-specific scores $\{\mathcal{S}_i(\cdot \mid s)\}_{i=1}^m$, corresponding to open-loop rule-based metrics in the Extended Predictive Driver Model Score. At inference time, trajectory selection is based on a weighted average of $m+1$ outputs: the policy probability and the $m$ predicted metric scores. The selected plan is thus the highest-ranked member of the fixed candidate set rather than a directly regressed continuous trajectory.

This design places the fixed trajectory vocabulary at the center of the method. A plausible implication is that ZTRS trades continuous action flexibility for exhaustive scoring and dense supervision across a large but finite decision set.

3. Offline reinforcement learning and Exhaustive Policy Optimization

ZTRS formulates planning as offline reinforcement learning over a discrete enumerable action space. The paper starts from the policy-gradient expression

$a_i$0

where $a_i$1 denotes the advantage function. Because the action space is finite and enumerable, the method rewrites optimization in a way that allows all candidate trajectories to contribute, rather than relying on a single sampled action (Li et al., 28 Oct 2025).

The resulting method is Exhaustive Policy Optimization (EPO), described as a variant of policy gradient tailored for offline data and enumerable actions. Its defining property is that it uses the likelihood form over the entire action set rather than the log-likelihood of one sampled action. The paper emphasizes that this produces denser supervision in trajectory selection, since every candidate trajectory contributes to the gradient signal.

The offline objective is defined over states from the dataset and actions from the fixed candidate set. In this setting, ZTRS does not interact with the environment during training. Instead, it evaluates candidate trajectories using offline reward computation. This is the mechanism by which the framework removes imitation learning while remaining practical on high-dimensional sensor inputs.

The paper’s comparison between “likelihood over all actions” and “log-likelihood on sampled action” is central to the method’s identity. The sampled-action variant underperforms, which the paper interprets as evidence that exhaustive supervision is advantageous when the planner must rank many discrete trajectories.

4. Reward construction, temporal consistency, and implementation

The primary reward signal in ZTRS is the Extended Predictive Driver Model Score (EPDMS), a composite metric reflecting safety, rule compliance, and progress (Li et al., 28 Oct 2025). The raw advantage is defined as

$a_i$2

where $a_i$3 is the EPDMS score and $a_i$4 is a temporal-consistency correction term. The previous action is

$a_i$5

and the correction term is

$a_i$6

Here, $a_i$7 is a constant, $a_i$8 denotes violation of Extended Comfort thresholds, and the indicator penalizes comfort-inconsistent changes between consecutive predicted trajectories. After computation, $a_i$9 is normalized to zero mean and unit variance.

For Navtest and Navhard, EPDMS is defined as

$s$0

The multiplicative penalty set $s$1 contains No-at-fault Collisions (NC), Drivable Area Compliance (DAC), Driving Direction Compliance (DDC), and Traffic Light Compliance (TLC). The weighted-average set $s$2 contains Time-to-Collision (TTC), Ego Progress (EP), Lane Keeping (LK), and History Comfort (HC). Extended Comfort (EC) is also used as a weighted metric to encourage temporally consistent driving.

The reported implementation details are specific. Training uses the Navtrain split, 24 NVIDIA A100 GPUs, 15 epochs, total batch size 528, learning rate $s$3, weight decay 0.0, and input resolution $s$4. The default action space contains 16,384 trajectories, each spanning 4 seconds at 10 Hz. Trajectory candidates are obtained by K-means clustering on nuPlan. The default image backbone is DD3D-pretrained V2-99, with ViT-L, pretrained from Depth-Anything, as an alternative. The temporal-consistency coefficient is set to $s$5.

5. Benchmarks and reported empirical performance

ZTRS is evaluated on three benchmarks: Navtest, Navhard, and HUGSIM (Li et al., 28 Oct 2025). Navtest is based on NAVSIM and is used for open-loop planning; the broader NAVSIM setup contains 103k training and 12k evaluation scenarios. Navhard extends evaluation to challenging NAVSIM scenarios with a two-stage procedure consisting of original NAVSIM open-loop evaluation and 3D Gaussian Splatting (3DGS)-based synthesized future scenarios; it contains 244 initial scenarios and 4164 synthetic scenarios. HUGSIM is a closed-loop benchmark with 3DGS-synthesized images, combining data from KITTI-360, Waymo, nuScenes, and Pandaset, for a total of 345 scenarios divided into 49 easy, 126 medium, 86 hard, and 84 extreme.

The main reported quantitative results are summarized below.

Benchmark	Setting	Reported result
Navhard	ZTRS (V2-99)	EPDMS 45.5
Navtest	ZTRS (V2-99)	EPDMS 85.3
Navtest	ZTRS (ViT-L)	EPDMS 86.2
HUGSIM	Zero-shot closed-loop	RC = 42.6, HD-Score = 28.9

On Navhard, ZTRS with V2-99 achieves 45.5 EPDMS and is reported as state-of-the-art. The paper describes this result as reflecting a strong balance of progress and comfort while maintaining high collision and compliance metrics. On Navtest, ZTRS reaches 85.3 with V2-99 and 86.2 with ViT-L. These results outperform Hydra-MDP++, but remain slightly behind DriveSuprim, which the paper attributes to more advanced data augmentation and scoring architecture.

On HUGSIM, ZTRS is evaluated in zero-shot closed-loop form, without adaptation, despite being trained only on offline real-world data and without imitation-learning pretraining. It achieves the best public-scenario Route Completion (RC) and HD-Score, with overall values 42.6 and 28.9 respectively. Relative to GTRS-Dense, the gains are 4.6% in RC and 0.3% in HD-Score.

The qualitative findings are aligned with these results. The paper reports that ZTRS generates trajectories resembling human driving behavior despite the absence of demonstrations and can handle safety-critical cases in HUGSIM, including an extreme overtaking scenario involving a parked car and an oncoming vehicle.

6. Ablations, limitations, and broader significance

The ablation studies isolate both the learning paradigm and the action-space design (Li et al., 28 Oct 2025). In the target-ablation study, the paper compares IL + RL with human target, IL + RL with $s$6 target, RL with likelihood over all actions (“ll”), and RL with log-likelihood on sampled action (“log-ll”). Several findings are explicit. Using the maximum-EPDMS trajectory as a single imitation target performs much worse than human imitation, which the paper attributes to the fact that many trajectories can have high EPDMS and a single best trajectory does not capture the structure of the action set. Optimizing likelihood over the full action space improves EPDMS substantially, but without the correction term it yields oscillation and poor EC. Adding the correction term $s$7 improves EC by 23.4%.

The action-space-size ablation studies training and inference spaces of 8192, 16384, and mixed configurations. The reported findings are not uniform across settings. Using the full action space at inference gives the best real-world results on Navtest and Navhard, but shrinking the inference action space can improve the simulated portion of Navhard. The paper gives a concrete example with V2-99: Train 16384 / Inference 8192 yields Navhard EPDMS 45.5, Train 16384 / Inference 16384 yields 43.4, and Train 8192 / Inference 8192 yields 43.0. This suggests that inference-time pruning can help in some synthetic environments.

The method’s limitations are also stated directly. ZTRS depends on a fixed discrete trajectory vocabulary and therefore does not directly output arbitrary continuous trajectories. Its performance depends on the coverage and quality of the candidate set. Its reward signals are derived from simulator-based or open-loop evaluation metrics, so its behavior is constrained by the fidelity of those metrics. Although it removes imitation learning, it still relies on carefully engineered driving metrics such as EPDMS and EC.

These limitations define the scope of the contribution rather than diminishing it. ZTRS demonstrates that end-to-end driving from high-dimensional sensor data need not be tied to imitation learning, provided that the action space is enumerable and the reward model is strong enough to supervise trajectory scoring. This suggests a distinct research direction in which objective safety, compliance, comfort, and progress metrics become the primary training signal for sensor-to-trajectory planning.