SEG-Parking: Offline RL for Autonomous Parking

• SEG-Parking is an end-to-end offline reinforcement learning framework that integrates structured state encoding with a conservative policy for autonomous parking.
• It fuses spatial, temporal, and goal information using a multi-branch encoder and utilizes a curated parking dataset to manage interactions, especially with opposite vehicles.
• CARLA simulator evaluations show that SEG-Parking achieves high success rates and robust generalization in both in-distribution and out-of-distribution parking scenarios.

SEG-Parking is an end-to-end offline reinforcement learning framework for autonomous parking designed for unstructured environments and interaction-rich scenarios, especially those involving an opposite vehicle (OV). Its reported formulation combines a specialized parking dataset, a goal-conditioned state encoder that maps fused perception information into a compact latent space, and an offline RL policy trained with a conservative regularizer that penalizes out-of-distribution actions. In closed-loop experiments in the high-fidelity CARLA simulator, the framework is reported to achieve the highest success rate among the compared baselines while maintaining robust generalization to out-of-distribution parking scenarios (Yang et al., 17 Sep 2025).

1. Problem setting and conceptual basis

SEG-Parking is positioned within a line of autonomous parking research concerned with unstructured environments, narrow maneuvering spaces, and dynamic interactions. The central difficulty is not merely geometric parking-slot entry, but interaction-aware decision making under clutter, restricted maneuver envelopes, and distribution shift between training and deployment conditions. The framework explicitly targets scenarios both without OV interference and with OV interactions, including cases in which the ego vehicle must yield or wait (Yang et al., 17 Sep 2025).

The work departs from planner-centric parking pipelines that first compute a maneuver and then refine it. For example, “Segmented Trajectory Optimization for Autonomous Parking in Unstructured Environments” preserves the maneuver phases of a high-level planner such as Hybrid A*, refines the initial path by SQP, allows curvature discontinuities at switching points, and reports convergence in two iterations with total computation times below $10$ ms in the studied scenarios (Yu et al., 7 Apr 2025). “Improving RRT for Automated Parking in Real-world Scenarios” instead uses a two-stage system consisting of a geometric in-slot planner and an enhanced RRT-based out-of-slot planner, reporting a $21\%$ parking path cost reduction and a $79.5\%$ computation-time reduction relative to its baseline, together with physical validation on a Porsche Cayenne (Pang et al., 2023). SEG-Parking addresses the same operational domain, but replaces explicit online search and trajectory refinement with an offline-learned policy (Yang et al., 17 Sep 2025).

A common misconception is that “end-to-end” parking systems necessarily operate on raw image streams. In SEG-Parking, the policy is end-to-end in the sense of learning a direct mapping from a fused parking state representation to control actions, but the state itself is structured and engineered rather than purely raw-visual (Yang et al., 17 Sep 2025). This distinction is important for interpreting both its strengths and its scope.

2. Scenario design, dataset, and task formulation

The training and evaluation data are constructed in CARLA simulator, specifically in Town04_Opt, with a focus on the central parking lot. The reported target slots are $\{S_{15}, S_{16}\}$ for in-distribution evaluation and $\{S_{11}, S_{12}\}$ for out-of-distribution generalization. For each scenario type, $100$ episodes are collected. The ego vehicle starts at the west side of the lot, the OV—when present—starts at the east side, and all non-target slots are filled with static obstacles represented as parked vehicles (Yang et al., 17 Sep 2025).

Scenario type	OV condition	Intended interaction pattern
Type i	Only EV, no OV interference	Basic parking
Type ii	OV present, low priority	OV yields to EV
Type iii	OV present, high priority	EV must yield or wait

Expert data are generated by a Hybrid A* planner combined with a finite-horizon LQR controller. To diversify the action distribution, the collection process adds $\epsilon$-greedy noise injection with $\epsilon = 0.2$. Each transition records $$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$ (Yang et al., 17 Sep 2025).

The state $s_t$ is composed of three main elements. The first is an obstacle distance sequence $21\%$0, defined as the last $21\%$1 LiDAR-derived vectors of obstacle distances, intended to capture short-term environment history. The second is a target pose $21\%$2, namely the vehicle’s current position and orientation relative to the goal slot, represented as $21\%$3. The third is ego motion $21\%$4, comprising instantaneous velocity and acceleration. The action $21\%$5 is a discretized 3-tuple of acceleration, steering angle, and gear, with $21\%$6 acceleration bins, $21\%$7 steering bins, and $21\%$8 gear states for forward and reverse (Yang et al., 17 Sep 2025).

This formulation makes the parking task goal-conditioned from the outset. It also makes interaction structure explicit: the learning problem is not restricted to static-slot alignment, but includes negotiation with an approaching opposite vehicle through the dataset design itself (Yang et al., 17 Sep 2025).

3. Goal-conditioned representation and policy architecture

The representational core of SEG-Parking is a goal-conditioned state encoder (GSE). Its purpose is to fuse spatial obstacle information, short temporal context, and goal-relative geometry into a compact latent variable $21\%$9 suitable for policy learning (Yang et al., 17 Sep 2025).

The temporal branch first differentiates $79.5\%$0 to obtain $79.5\%$1, thereby emphasizing change in the obstacle-distance history. The differentiated sequence is then processed by an MLP and an LSTM to produce a temporal feature $79.5\%$2. In parallel, the spatial branch processes the current distance vector $79.5\%$3 with another MLP to obtain $79.5\%$4. A goal branch maps the relative target pose $79.5\%$5 through a feedforward projection to form $79.5\%$6 (Yang et al., 17 Sep 2025).

Goal conditioning is implemented through multi-head cross-attention: $79.5\%$7 with $79.5\%$8 serving as the query and the concatenated temporal and spatial features serving as memory. The output is the latent state $79.5\%$9, described as a goal-conditioned, context-integrated, compact representation (Yang et al., 17 Sep 2025).

The policy network $\{S_{15}, S_{16}\}$0 is an MLP with ReLU nonlinearities and a final $\{S_{15}, S_{16}\}$1 layer, producing the action prediction $\{S_{15}, S_{16}\}$2. Value estimation is performed by twin Q-networks $\{S_{15}, S_{16}\}$3 and $\{S_{15}, S_{16}\}$4, also implemented as MLPs. The twin-critic design is used to reduce overestimation bias during offline RL (Yang et al., 17 Sep 2025).

Architecturally, SEG-Parking therefore remains structured even while being described as end-to-end. The encoder imposes a decomposition into temporal, spatial, and goal branches, and the policy is learned over a compact latent space rather than over full trajectory optimization variables. This suggests a compromise between pure representation learning and classical modular parking pipelines (Yang et al., 17 Sep 2025).

4. Learning pipeline and objective functions

Training proceeds in two stages: behavior-cloning pretraining of the encoder, followed by offline RL fine-tuning with conservative regularization. In the pretraining stage, an action head $\{S_{15}, S_{16}\}$5 is attached to the encoder $\{S_{15}, S_{16}\}$6, and the model is trained using the behavior-cloning objective

$\{S_{15}, S_{16}\}$7

The encoder is not frozen afterward; it continues to be fine-tuned during RL optimization (Yang et al., 17 Sep 2025).

The offline RL stage uses twin-critic temporal-difference learning. The reported TD objective is

$\{S_{15}, S_{16}\}$8

where $\{S_{15}, S_{16}\}$9 denotes the nearest discrete action to $\{S_{11}, S_{12}\}$0 under a Wolpertinger-style projection, which is required because the policy output is mapped back into the discretized action space (Yang et al., 17 Sep 2025).

The distinctive component is the conservative regularizer: $\{S_{11}, S_{12}\}$1 Here, $\{S_{11}, S_{12}\}$2 controls regularization strength and $\{S_{11}, S_{12}\}$3 controls the softness of the log-sum-exp term. The effect is to penalize Q-value inflation on actions that are not well supported by the offline dataset, thereby discouraging risky out-of-distribution actions at deployment (Yang et al., 17 Sep 2025).

Policy improvement is then expressed as

$\{S_{11}, S_{12}\}$4

The reward is decomposed as

$\{S_{11}, S_{12}\}$5

The term $\{S_{11}, S_{12}\}$6 is a terminal goal bonus when the vehicle is parked within the prescribed tolerances. The term $\{S_{11}, S_{12}\}$7 is an exponentially decaying progress reward as distance to the goal decreases. The term $\{S_{11}, S_{12}\}$8 penalizes collisions. The term $\{S_{11}, S_{12}\}$9 penalizes unsafe behavior when the agent’s action deviates too far from expert action, specifically when $100$0 (Yang et al., 17 Sep 2025).

A second misconception is that offline RL in parking is equivalent to imitation learning on a fixed log. SEG-Parking is not limited to supervised action regression. It first learns a behavior prior through BC, but subsequently optimizes a value-based objective with explicit conservatism and policy improvement over the offline support (Yang et al., 17 Sep 2025).

5. Closed-loop evaluation and reported performance

The implementation is reported on Ubuntu 22.04 with an NVIDIA A100 GPU. Encoder pretraining runs for $100$1 epochs, and RL fine-tuning runs for $100$2 epochs with batch size $100$3 and the Adam optimizer. The listed hyperparameters include $100$4, $100$5, and $100$6 (Yang et al., 17 Sep 2025).

The baselines are SAC, SAC-$100$7, Behavior Cloning, and TD3-BC. Evaluation uses TSR (Target Success Rate), TFR (Target Failure Rate), CR (Collision Rate), TR (Timeout Rate), and, for non-collision and non-timeout completions, APE (Average Position Error), AOE (Average Orientation Error), and APT (Average Parking Time) (Yang et al., 17 Sep 2025).

In the in-distribution slots, SEG-Parking is reported with TSR values in the range $100$8–$100$9, CR in the range $\epsilon$0–$\epsilon$1, average position error around $\epsilon$2 m, average orientation error around $\epsilon$3, and average parking time around $\epsilon$4–$\epsilon$5 s. Over the same set of experiments, BC is reported at $\epsilon$6–$\epsilon$7 TSR with $\epsilon$8–$\epsilon$9 collision, TD3-BC at $\epsilon = 0.2$0–$\epsilon = 0.2$1 TSR, and SAC-$\epsilon = 0.2$2 at $\epsilon = 0.2$3–$\epsilon = 0.2$4 TSR with more aggressive behavior (Yang et al., 17 Sep 2025).

The out-of-distribution results are the most consequential. On unseen slots, SEG-Parking is reported to maintain TSR in the range $\epsilon = 0.2$5–$\epsilon = 0.2$6 with collision rates of $\epsilon = 0.2$7–$\epsilon = 0.2$8. By contrast, BC drops to $\epsilon = 0.2$9–$$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$0 TSR with $$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$1–$$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$2 collision, TD3-BC reaches only $$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$3–$$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$4 TSR with $$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$5–$$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$6 collision, and SAC-$$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$7 degrades to $$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$8–$$(s_t, a_t, r_t, s_{t+1}, \mathbb{I}\{\text{parked}\})$$9 TSR with $s_t$0–$s_t$1 collision (Yang et al., 17 Sep 2025).

The reported closed-loop visualizations further show interaction-aware behavior. In the described type-iii setting, the ego vehicle waits for the OV when required and proceeds smoothly into the slot once permitted. This matters because the framework’s claimed generalization is not limited to slot geometry; it extends to tactical interaction behavior learned from the offline dataset (Yang et al., 17 Sep 2025).

6. Relation to adjacent parking paradigms and current limits

SEG-Parking belongs to a broader shift from modular parking stacks toward learned planning and control, but it differs materially from several neighboring lines of work. “Automated Parking Trajectory Generation Using Deep Reinforcement Learning” formulates parking as an MDP with continuous steering and throttle and trains a Soft Actor-Critic policy in Gazebo; it reports high success rates, robust handling of dynamic obstacles, and faster or comparable planning times than Hybrid A* across three cases, but its method is online DRL rather than offline conservative RL (Zhang et al., 29 Apr 2025). “ParkingWorld” uses correction-in-the-loop sample-efficient RL in a 3D Gaussian Splatting simulator and reports PSR $s_t$2 in simulation and $s_t$3 in real-world experiments, emphasizing human corrective interventions and sim-to-real transfer (Yu et al., 24 May 2026). “ParkingTransformer” introduces an LLM-enhanced end-to-end planner with multi-view perception, 3D positional encoding, and fixed-window history, and reports a driving score of $s_t$4 in CARLA and an average real-world success rate of $s_t$5 (Wu et al., 12 Jun 2026). “MultiPark” takes a multimodal imitation-learning route using next-segment prediction, factorized parking queries, and outcome-oriented losses, reporting a $s_t$6 success rate, $s_t$7 collision rate on the HP5 dataset, and approximately $s_t$8 ms per plan on production-car hardware (Zheng et al., 15 Aug 2025).

Within that landscape, SEG-Parking is distinguished by three features: a parking-specific offline dataset with explicit OV interaction regimes, a goal-conditioned latent encoder over structured parking state, and conservative Q-regularization to suppress unsupported actions (Yang et al., 17 Sep 2025). A related misconception is that all learned parking systems solve the same problem with interchangeable representations. In fact, SEG-Parking does not use raw multi-view images, dense BEV trajectory queries, or multimodal segment decoders; it uses LiDAR-derived obstacle-distance histories, relative goal pose, and ego motion (Yang et al., 17 Sep 2025).

Its present limitations are also clear from the reported setup. The dataset is constructed in CARLA Town04_Opt around one central parking lot, the target slots are explicitly enumerated, the action space is discretized, and the experiments are closed-loop simulations rather than physical-vehicle trials (Yang et al., 17 Sep 2025). This suggests that the framework’s main contribution lies in showing that conservative offline RL can remain both high-performing and interaction-aware under controlled parking distribution shifts, rather than in establishing full real-world deployment readiness.