SunnyParking: Multi-Shot Trajectory Generation and Motion State Awareness for Human-like Parking

arXiv:2602.21682 · cs.RO · Submitted 2026-02-25 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "SunnyParking: Multi-Shot Trajectory Generation and Motion State Awareness for Human-like Parking".

Dev: Here's the gist of the paper. Autonomous parking fundamentally differs from onroad driving due to its frequent direction changes and complex maneuvering requirements <ref:2602.21682#pg2>.

Rosa: This paper proposes SunnyParking, a novel dual-branch E2E architecture that achieves motion state awareness by jointly predicting spatial trajectories and discrete motion state sequences.

Dev: Additionally, it introduces a Fourier feature-based representation of target parking slots to overcome the resolution limitations of traditional bird’s-eye view (BEV) approaches.

Rosa: Experimental results demonstrate that our framework generates more robust and human-like trajectories in complex multi-shot parking scenarios, while significantly improving gear-shift point localization accuracy compared to state-of-the-art methods.

Paper summary: Rosa: So looking at SunnyParking: Multi-Shot Trajectory Generation and Motion State Awareness for Human-like Parking, the core idea is this dual task parallel prediction framework that jointly predicts spatial trajectories and discrete motion state sequences.

Dev: The authors of this paper claim they achieved motion state awareness by jointly predicting those spatial trajectories and those discrete motion state sequences.

Taro: It seems to be a way of solving the dimensionality deficiency that existing end to end planning methods have, which is when they simplify the parking task into a geometric path regression problem.

Rosa: They also introduced Fourier feature mapping for target parking slots, which helps overcome resolution limitations from traditional bird's eye view approaches.

Dev: The implication is that by fusing spatial trajectory prediction with motion state prediction and using this Fourier representation of targets, we get trajectories that are more physically feasible and look more human like in complex scenarios.

Taro: It suggests that for autonomous systems operating in complex maneuvers, just predicting where to go isn't enough; you have to model what the vehicle is actually doing kinematically.

Rosa: And it means that for our users, it translates into a parking system that handles those frequent direction changes and complex maneuvers much better than current methods.

Dev: So in simple terms, SunnyParking shows that adding a motion state branch significantly enhances the model's ability to predict those motion state transition points within multi shot parking trajectories.

Taro: It confirms that understanding the motion state is a prerequisite for accurate complex trajectory generation when you're trying to follow expert behaviors.

Rosa: And they also show that their method sensitively captures distinct human driver parking styles, which is something we need to consider when training these systems.

Conclusion: Rosa: So, we're wrapping up our look at SunnyParking by looking at what this whole paper actually means for autonomous parking systems out in the real world.

Dev: It boils down to them proposing this new dual task parallel prediction framework that handles both where the car needs to go and what motion state it should be in.

Rosa: Yeah, they’re focusing on making these predictions together, not separately, which is a big change for how you'd model a complex maneuver.

Dev: And they used something called Fourier feature mapping to represent the parking slots themselves, trying to fix the problem where traditional bird's eye view methods just can't see the target slot clearly enough.

Rosa: It seems like their main win is that they get trajectories that actually look more human-like when those parking situations get really messy with multiple shots.

Dev: They claim a sixty nine point six percent error reduction in planning compared to the best methods out there currently available, which is solid if it holds up in testing.

Rosa: The authors also pointed out that modeling the motion states explicitly makes a huge difference in reducing the distance error, showing that you really need to understand what's happening kinematically for good path generation.

Dev: That means for someone actually driving or controlling a robot, it suggests that knowing if you’re supposed to be accelerating or braking at this exact moment is just as important as knowing the precise coordinates of the next waypoint.

Rosa: It really moves us beyond just drawing a line between points; it’s about modeling the actual driving style of a person trying to park.

Dev: It also shows they can handle those tricky gear shifts and accurately predict how many shifts are needed even in those four shot scenarios, which is crucial for safety cases.

Rosa: So, the big picture here is that by adding this motion state awareness, autonomous parking gets much better at handling the frequent direction changes and complex maneuvers we see every day.

Dev: But we need to keep asking about how robust this model is when things go completely off script in those real-world trials.

Chubu University of Technology, Kasugai, Japan

cs.RO

Submitted: 2026-02-25

Updated: 2026-11-12

Comments: Accepted for publication in IEEE Access

Journal ref: IEEE Access, 2026

DOI: 10.1109/ACCESS.2026.3752444

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 79/100

The gist: Autonomous parking fundamentally differs from onroad driving due to its frequent direction changes and complex maneuvering requirements <ref:2602.21682#pg2>.

Key concepts

Dual-Branch E2E Architecture
This is the core design where the system has two separate prediction paths: one branch focuses on predicting the spatial path (waypoints) of the car, and another branch focuses on predicting discrete motion states (like braking or shifting). They work together to create a complete plan.
Fourier Feature Mapping
This technique is used to represent the target parking slot information. Instead of using a fixed grid like traditional bird's-eye view methods, this method maps the position into a higher-dimensional feature vector. This allows the system to capture fine details and achieve better precision for locating parking spots.
Motion State Awareness
This refers to the model's ability to predict *how* the vehicle is moving at any given moment, not just where it is going. By predicting discrete motion states alongside trajectories, the system gains a deeper understanding of complex maneuvers required for human-like parking.

Terminology

Summary

Autonomous parking fundamentally differs from onroad driving due to its frequent direction changes and complex maneuvering requirements <ref:2602. p. 021682#pg2>. This paper proposes SunnyParking, a novel dual-branch E2E architecture that achieves motion state awareness by jointly predicting spatial trajectories and discrete motion state sequences. p. 0 p. 0 Additionally, it introduces a Fourier feature-based representation of target parking slots to overcome the resolution limitations of traditional bird’s-eye view (BEV) approaches. p. 0 p. 0 Experimental results demonstrate that our framework generates more robust and human-like trajectories in complex multi-shot parking scenarios, while significantly improving gear-shift point localization accuracy compared to state-of-the-art methods. p. 0 p. 0

Problem Definition and Dataset Construction

The core idea is to train a neural network to map sensor inputs to future vehicle trajectories based on expert behaviors. The dataset consists of M expert-driven trajectories, where each trajectory i is a time-ordered sequence of states with a total length of Ni. p. 2 The available data for any given time step j includes the RGB image I, parking trajectory waypoint T, parking motion states M, and target parking slot position S. p. 2 The network Nθ takes the current images and goal as input and outputs predictions for the future trajectory and motion states structured to match ground truth targets. p. 3

Target Representation Enhancement

The authors address the resolution limitations of traditional BEV approaches by designing a cross-domain target fusion mechanism that represents the target slot using Fourier feature mapping. This approach is not constrained by fixed BEV resolution or performance degradation at the grid’s edges, overcoming a key limitation of traditional BEV-based schemes. p. 1 The position encoding uses a function γ(v) which maps an input low-dimensional coordinate vector v to a higher-dimensional feature vector. p. 1 By setting the parameter L = 12, the theoretical resolution is around 1. p. 32 cm, and the practical effective precision of our method reaches 1 − 2 mm. p. 3

Dual-Branch Prediction Architecture

The proposed architecture is called YUKUI, which resembles a Y-shaped Unified Kinematic and Uncertainty-aware Imitation planner. It consists of two main branches: the Trajectory Branch and the Motion State Branch. p. 4 The Trajectory Branch uses a Transformer decoder for predicting waypoints, where target-aware features Fenchanced serve as Key and Value, while a set of N learnable queries serve as the Query input. p. 4 The Motion State Branch predicts the vehicle motion state using a two-stage query fusion mechanism. p. 4 This involves refining Motion Query embeddings by attending to hidden layer features from the trajectory branch before using a lighter Transformer Decoder. p. 4

Loss Function Design and Evaluation Metrics

The network’s parameters θ are optimized by minimizing a composite loss function Ltotal, which is a weighted sum of two component losses: Ltotal = λwaypointsLwaypoints + λmotionLmotion. The waypoint loss Lwaypoints is the mean squared error (MSE) between T and Tˆ. p. 4 The motion loss Lmotion is combined with a Mean Cross-Entropy (MCE) loss λCE between the probability distributions in Mˆ ′ and the one-hot labels in M′. p. 4 A GT-awareness smoothness loss λsmooth is implemented as a masked L1 loss that deactivates calculation around ground truth gear-shift locations to enforce smoothness only in stable regions. p. 4 Evaluation metrics include L2 Distance error, Fourier descriptor distance error, and Hausdorff distance error for trajectory evaluation. p. 5 Motion State Accuracy is defined as the percentage of timesteps where the dominant motion state is correctly predicted. p. 5

Performance and Ablation Insights

Experimental results demonstrate that our method achieves a 69.6% error reduction in trajectory planning compared to state-of-the-art approaches. p. 1 p. 1 The ablation studies show that explicitly modeling motion states drastically reduces the L2 Distance error, confirming that understanding the motion state is a prerequisite for accurate complex trajectory generation. p. 6 p. 6 The method maintains accurate positioning between shift points and precisely predicts the number of shifts even within the most complex 4-shot scenarios. p. 6 p. 6 Qualitative visualization shows that our method proves significantly more robust than baseline methods, maintaining high accuracy in predicting gear-shift frequency and location while ensuring trajectory smoothness. p. 7 p. 7

Conclusion

In this study, we propose an end-to-end dual-task parallel trajectory prediction framework. By introducing an additional motion state branch, the model significantly enhances its ability to predict the motion state transition points in multi-shot parking trajectories. p. 7 We employ Fourier feature mapping to encode target point into query vectors, which interact with sensor features in the BEV to achieve cross-domain information fusion. p. 7 Experimental results demonstrate that our dual-branch model not only exhibits a superior trajectory handling capability in complex parking scenarios but also sensitively captures distinct human driver parking styles. p. 7 The future work will focus on extending the study to more complex real-world scenarios, prioritizing enhancements in the model’s ability to predict interactions with dynamic traffic participants and deepening its semantic understanding of complex environments. p. 7">

<ref:2602.21682#pg2>Autonomous parking fundamentally differs from onroad driving due to its frequent direction changes and complex maneuvering requirements. p. 0 This paper proposes SunnyParking, a novel dual-branch E2E architecture that achieves motion state awareness by jointly predicting spatial trajectories and discrete motion state sequences. Additionally, it introduces a Fourier feature-based representation of target parking slots to overcome the resolution limitations of traditional bird’s-eye view (BEV) approaches. Experimental results demonstrate that our framework generates more robust and human-like trajectories in complex multi-shot parking scenarios, while significantly improving gear-shift point localization accuracy compared to state-of-the-art methods.

Experimental results demonstrate that our method achieves a 69.6% error reduction in trajectory planning compared to state-of-the-art approaches. The ablation studies show that explicitly modeling motion states drastically reduces the L2 Distance error, confirming that understanding the motion state is a prerequisite for accurate complex trajectory generation. The method maintains accurate positioning between shift points and precisely predicts the number of shifts even within the most complex 4-shot scenarios. Qualitative visualization shows that our method proves significantly more robust than baseline methods, maintaining high accuracy in predicting gear-shift frequency and location while ensuring trajectory smoothness.">

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Dual-branch architecture for joint prediction allows a model to simultaneously output a high-precision spatial trajectory and a strictly aligned sequence of discrete motion states. This enables more robust planning by explicitly modeling both path geometry and required vehicle commands like forward/reverse transitions.

  2. Fourier feature mapping overcomes resolution limitations by allowing the target slot to be conditioned on precise, high-fidelity goal information, preserving sub-centimeter accuracy independent of the fixed BEV grid resolution. This leads to more reliable attention modulation during multi-shot maneuvers.

  3. The cross-domain target fusion module utilizes a set of learnable spatial queries that are modulated by the Fourier encoded target, allowing the planner to extract the relevant information from the BEV feature map in a context-aware manner. This ensures attention remains strictly aligned with the immediate planning intent throughout the long-term maneuver.

  4. The YUKUI planner serializes both trajectory waypoints and motion states, converting prediction into a token prediction problem. This allows for a one-to-one correspondence at each time step between predicted waypoints and motion state sequences, ensuring kinematic feasibility.

  5. The loss function incorporates a GT-awareness smoothness loss that penalize[s] the network for predicting correct, clear gear-shift points by deactivating the loss calculation with a tolerance window surrounding ground truth locations. This specifically addresses the motion state shaking in hesitant areas.

Sources

Related papers