U-Parking: Distributed UWB-Assisted Autonomous Parking System with Robust Localization and Intelligent Planning

This paper presents U-Parking, a distributed autonomous parking system that combines Ultra-Wideband-assisted robust localization with Large Language Model-driven intelligent planning to achieve reliable automated parking in challenging indoor environments, as demonstrated on real vehicles.

The Gist

Imagine you are trying to park your car in a massive, confusing indoor garage. The lights are dim, there are pillars everywhere, and your GPS (which usually works great outside) is completely useless inside. You try to back in, but you keep hitting a wall, or you spin in circles because you don't know exactly where you are.

This is the problem U-Parking solves. It's a new system that lets cars park themselves automatically, even in the trickiest indoor garages, by combining three smart technologies: a "super-sense" for location, a "smart brain" for planning, and a "steady hand" for driving.

Here is how it works, broken down into simple concepts:

1. The "Super-Sense": Finding the Car (Robust Localization)

The Problem: Inside a garage, radio signals bounce off walls and pillars (like an echo in a cave). This confuses standard sensors, making the car think it's in one spot when it's actually three feet away. This causes the car to "jump" around on the screen or drift off course.

The U-Parking Solution:
Think of the car wearing a smart watch (a UWB tag) and the garage having smart lighthouses (UWB anchors) on the ceiling.

• The Magic: Instead of just listening to one signal, the system uses a special "fusion" technique. It combines the radio signals with the car's own motion sensors (like the gyroscope in your phone).
• The Analogy: Imagine you are walking in the dark. You can't see, but you have a friend shouting directions (the UWB signal). Sometimes your friend shouts the wrong direction because of an echo. But, you also know exactly how many steps you've taken and which way you turned (the motion sensor). The system acts like a smart referee that says, "Okay, the signal says you're here, but your steps say you're there. Let's trust your steps a bit more right now to smooth out the error." This stops the car from getting confused by "ghost" signals.

2. The "Smart Brain": Choosing Where to Park (LLM-Assisted Planning)

The Problem: Traditional parking systems are like a robot following a strict map. If the map says "Spot 10 is open," it tries to go there, even if the path is blocked by a temporary obstacle or the signal is bad. It wastes time checking every single empty spot.

The U-Parking Solution:
This system uses a Large Language Model (LLM)—the same kind of AI that writes essays or chats with you—as a Parking Manager.

• The Magic: Before the car even starts moving, the "Manager" looks at the whole garage. It doesn't just look at a map; it understands context. It thinks: "Hey, Spot 10 is open, but the signal there is weak and there's a pillar blocking the way. Spot 25 is also open, the signal is strong, and the path is clear. Let's pick Spot 25."
• The Analogy: Imagine you are looking for a seat in a crowded movie theater. A normal robot would check every row one by one. The U-Parking AI is like a concierge who knows the theater layout, knows which rows have bad views, and instantly points you to the best seat, saving you from walking down every aisle. This makes the planning much faster and smarter.

3. The "Steady Hand": Driving the Car (Robust Control)

The Problem: Even with a good map and a good location, if the car gets a sudden "jolt" of bad data, it might panic, slam on the brakes, or jerk the steering wheel, causing a crash or a shaky ride.

The U-Parking Solution:
The car uses a Model Predictive Controller (MPC). Think of this as a defensive driver.

• The Magic: If the "Super-Sense" (Step 1) suddenly gets a weird reading, the "Steady Hand" doesn't panic. It says, "Okay, the data looks a bit shaky right now. I'm going to ignore the tiny errors and focus on keeping the car smooth and steady." It adjusts its driving style based on how confident it is in its location.
• The Analogy: Imagine driving on a bumpy road. A nervous driver would jerk the wheel every time the car hits a bump. A defensive driver (U-Parking) knows the road is bumpy, so they hold the wheel steady and absorb the bumps, ensuring the ride remains smooth for the passengers.

The Result: A Real-World Demo

The team actually built this system and tested it with a real car in a real indoor parking lot.

• Without U-Parking: The car would get confused, drift, or fail to park.
• With U-Parking: The car successfully found a spot, navigated around obstacles, and parked itself smoothly, even when the radio signals were bouncing off walls.

In a nutshell: U-Parking is like giving a self-driving car a super-accurate internal compass, a wise parking manager, and a calm, experienced driver, all working together to solve the headache of indoor parking.

Technical Summary

Based on the provided paper, here is a detailed technical summary of U-Parking:

1. Problem Statement

The paper addresses the challenges of autonomous parking in complex indoor environments (e.g., multi-story parking garages). Key difficulties include:

• Localization Instability: Indoor environments suffer from occlusions, multipath effects, and frequent transitions between Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) conditions. While Ultra-Wideband (UWB) offers centimeter-level accuracy, it is prone to "positioning jumps" and drift in these conditions, leading to control oscillations or parking failures.
• Planning Inefficiency: Traditional Simultaneous Localization and Mapping (SLAM) methods rely on high-precision prior maps and are sensitive to environmental changes. Conventional search-based planners (like A*) suffer from large search spaces and low efficiency in large-scale facilities, lacking global decision-making capabilities regarding parking space availability and signal quality.

2. Methodology

U-Parking is a distributed, collaborative system built on a ROS2 framework that integrates vehicle-side perception with server-side intelligence. The architecture consists of three core components:

A. Intelligent Planning (Server-Side)

• LLM-Assisted Decision Making: Instead of relying solely on classical search algorithms, the system employs a Large Language Model (LLM) (specifically LLaMA-3.3-70B) on the parking-lot server.
• Workflow:
  1. The server preprocesses the parking lot map to extract obstacle geometry and candidate nodes.
  2. Upon receiving a request, the LLM analyzes structured inputs (obstacle locations, UWB anchor deployment, vehicle pose, and real-time space availability).
  3. The LLM prioritizes parking spaces with favorable UWB signal quality and consecutive availability, outputting a target space ID and a set of key waypoints.
  4. These waypoints guide an Improved A* planner, significantly reducing the search space and unnecessary grid expansions.
• Trajectory Generation: The resulting path is smoothed using piecewise B-spline interpolation and Bézier curves, with the final parking maneuver generated via a curve-line geometric method.

B. Layered Fusion Positioning (Vehicle-Side)

To mitigate UWB errors caused by NLOS and multipath effects, the system uses a three-layer hierarchical fusion localization strategy:

1. Preprocessing: Raw UWB ranging data is filtered (1D Kalman Filter) to suppress noise.
2. Initial Estimation: A smoothed 2D position estimate is computed from the filtered ranging data.
3. Adaptive Optimization (IAEKF): An Improved Adaptive Extended Kalman Filter (IAEKF) fuses UWB data with IMU (Inertial Measurement Unit) motion states.
   • Mechanism: The filter calculates the innovation residual ($r_k$). If the residual exceeds a tuned threshold (indicating a potential NLOS jump), the filter dynamically reduces the weighting of the UWB measurement, relying more on the IMU propagation to maintain stability.

C. Robust Trajectory Tracking Control

• The system uses a Model Predictive Control (MPC) controller designed to handle positioning uncertainty.
• Adaptive Cost Weights: The controller monitors localization reliability. When severe UWB-induced jumps are detected, it down-weights instantaneous position errors and increases penalties for excessive velocity and steering variations. This prevents aggressive control inputs and vehicle oscillations.

3. Key Contributions

1. Hierarchical UWB Fusion Localization: A novel strategy that significantly improves robustness during LOS/NLOS transitions, mitigating the meter-level drift common in conventional UWB systems.
2. LLM-Assisted Global Planning: Introduces a server-side LLM mechanism that guides high-level decision-making (space selection and waypoint generation), effectively constraining the path search space and improving planning efficiency compared to standard A*.
3. Collaborative Vehicle-Parking Lot Architecture: Enables unified global positioning and decision-making that transcends locally perceived maps, leveraging the infrastructure (anchors) for better reliability.
4. Robust Control under Uncertainty: An adaptive MPC controller that specifically accounts for localization noise to ensure smooth trajectory tracking.

4. Experimental Results

• Setup: Experiments were conducted in a 30.2m × 37.9m indoor area using a SCOUT 2.0 differential-drive vehicle, HR-RTLS1 UWB platform, and WHEELTEC N100 IMU.
• Localization Performance: As shown in Figure 2(b) and Table I, the proposed Integrated (UWB+IAEKF) method outperformed baselines (UWB alone, UWB+EKF, UWB+UKF).
  • Mean Euclidean Error: Reduced from 0.240m (UWB) to 0.138m (UWB+IAEKF).
  • Max Error: Reduced from 2.354m to 0.653m.
  • Dynamic Time Warping (DTW): Improved from 0.284m to 0.169m.
• Tracking Performance: The system achieved stable parking with a mean error of 0.118m for the full integrated system. Occasional peak errors (up to 0.5m) were observed but were rare, short-lived, and did not compromise overall parking smoothness.
• Real-World Validation: The system was successfully demonstrated on real vehicles, completing autonomous parking tasks in a challenging indoor environment with frequent signal switching.

5. Significance

U-Parking represents a significant step forward in indoor autonomous driving by bridging the gap between infrastructure-based sensing and vehicle control.

• Reliability: It solves the critical issue of UWB instability in NLOS conditions, making automated parking viable in real-world, cluttered indoor garages without requiring expensive LiDAR or perfect map pre-scanning.
• Scalability: The distributed architecture (Server-Client) allows for centralized management of large parking facilities, optimizing space usage and traffic flow.
• AI Integration: It demonstrates the practical utility of Large Language Models in robotics, moving beyond simple chat interfaces to complex spatial reasoning and path planning optimization.

In conclusion, U-Parking validates that combining LLM-guided planning with robust sensor fusion and adaptive control creates a highly reliable system for autonomous parking in complex, dynamic indoor environments.