Unified Linear Parametric Map Modeling and Perception-aware Trajectory Planning for Mobile Robotics

Hongyu Nie, Xingyu Li, Xu Liu, Zhaotong Tan, Sen Mei, Wenbo SuarXiv 2025 · DOI · arXiv

Abstract

Autonomous navigation in mobile robotics requires a delicate balance between environmental perception and trajectory planning. Traditional methods often suffer from high computational costs, resolution sensitivity, and a lack of differentiability in mapping representations. This paper presents a unified framework utilizing Random Mapping and Random Projection (RMRP) to transform environmental mapping into a continuous, differentiable linear model. By leveraging analytical gradients and predictive scene completion, the proposed framework enables proactive, perception-aware planning for both Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs), significantly improving safety and efficiency in complex, cluttered environments.

Introduction

Mobile robots, whether flying or ground-based, face significant challenges in navigating unknown environments. UAVs often encounter sensor blind spots that lead to late reactions to obstacles, while UGVs can fall into negative obstacles like pits because standard occupancy maps fail to represent terrain depth accurately.

Overview of perception-aware planning for UAVs and UGVs.

Historically, these issues have been addressed by separate mapping and planning modules that rely on discrete structures like Occupancy Grid Maps or Euclidean Signed Distance Fields (ESDFs). However, these structures are resolution-dependent and require computationally expensive interpolation to generate gradients for trajectory optimization. This research introduces a paradigm shift by treating mapping as a regression task, resulting in a "resolution-invariant" map that provides exact mathematical gradients for smoother movement.

Methodology: The RMRP Framework

The core of the proposed system is the Random Mapping and Random Projection (RMRP) module, which works in tandem with the Robust Perception-Aware Trajectory Planner (RPATR).

The proposed system architecture integrating mapping and planning.

1. Linear Parametric Modeling

Instead of storing a massive grid of voxels, RMRP learns a linear parametric model from raw point cloud data:

  • Random Mapping (RM): Raw data is projected into a high-dimensional feature space where obstacles and free space become linearly separable.
  • Random Projection (RP): To maintain efficiency, this high-dimensional data is compressed while preserving critical geometric structures. This is supported by the authors' Residual Energy Preservation Theorem, which mathematically guarantees that the compressed model retains the essential shape of the environment (e.g., wall edges or pit boundaries).

2. Analytical Gradients and Scene Completion

Because the resulting map is a continuous mathematical function, the trajectory optimizer can calculate closed-form gradients directly. This eliminates the noise and latency associated with finite-difference methods used in traditional planners. Furthermore, the model can be trained on historical 3D data to perform predictive scene completion, allowing the robot to "fill in" occluded areas behind obstacles and plan proactive paths through potential blind spots.

Experimental Evaluation

The framework was rigorously tested in both simulation (Gazebo) and real-world scenarios.

  • UAV Testing: A quadrotor equipped with a Livox Mid360 LiDAR and an NVIDIA Jetson Xavier NX was flown in a 15m x 25m indoor arena. The RMRP approach consistently outperformed state-of-the-art (SOTA) planners in terms of path smoothness and total flight time.
  • UGV Testing: A SCOUT 2.0 ground robot with a Velodyne VLP-16 LiDAR navigated outdoor terrain. While standard planners like Fast-Planner often failed to detect deep pits (treating them as safe empty space), the RMRP-based planner achieved zero pit collisions.

Key Insights and Results

  • Safety Gains: The perception-aware strategy increased safety margins by over 600% in critical "blind corner" scenarios by anticipating hidden obstacles.
  • Efficiency: The parametric representation is significantly more memory-efficient than Bayesian filtering, as it only requires storing a small set of model parameters ($\beta$).
  • Smoothness: Qualitative analysis shows that the analytical gradients lead to trajectories with near-zero numerical error, providing much more stable flight and drive paths compared to traditional discrete maps.

Conclusion

By unifying mapping and planning through a linear parametric lens, this research provides a lightweight yet powerful solution for mobile robot navigation. The RMRP framework effectively bridges the gap between perception and action, allowing robots to navigate complex terrains and cluttered spaces with unprecedented safety and mathematical precision. Future work will focus on extending this framework to handle high-speed dynamic obstacles and multi-robot collaborative mapping.

Summary updated

BibTeX
@misc{nie2025unified,
  author = {Hongyu Nie and Xingyu Li and Xu Liu and Zhaotong Tan and Sen Mei and Wenbo Su},
  title = {{Unified Linear Parametric Map Modeling and Perception-aware Trajectory Planning for Mobile Robotics}},
  year = {2025},
  eprint = {2507.09340},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.09340},
}
Figure