SHIFT Planner: Speedy Hybrid Iterative Field and Segmented Trajectory Optimization with IKD-tree for Uniform Lightweight Coverage

TL;DR

The SHIFT planner addresses the need for uniform, semantic-aware coverage in complex 3D environments by integrating terrain-aware surface extraction, direct landmark point planning, and a Radiant Field-Informed Coverage Planning (RFICP) that modulates speed via a Gaussian diffusion field. Real-time obstacle avoidance is achieved through IKD-SWOpt, which combines an IKD-tree distance field-guided A* search with non-compliant segment identification and sliding-window optimization, followed by B-spline smoothing to preserve continuity. Extensive simulations and hardware experiments demonstrate superior coverage completeness and uniformity, along with efficient and responsive local trajectory refinement compared to state-of-the-art baselines. The approach promises robust performance in dynamic, semantically rich environments and lays groundwork for multi-robot extensions for large-scale uniform coverage tasks.

Abstract

This paper introduces a comprehensive planning and navigation framework that address these limitations by integrating semantic mapping, adaptive coverage planning, dynamic obstacle avoidance and precise trajectory tracking. Our framework begins by generating panoptic occupancy local semantic maps and accurate localization information from data aligned between a monocular camera, IMU, and GPS. This information is combined with input terrain point clouds or preloaded terrain information to initialize the planning process. We propose the Radiant Field-Informed Coverage Planning algorithm, which utilizes a diffusion field model to dynamically adjust the robot's coverage trajectory and speed based on environmental attributes such as dirtiness and dryness. By modeling the spatial influence of the robot's actions using a Gaussian field, ensures a speed-optimized, uniform coverage trajectory while adapting to varying environmental conditions.

Figures (6)

• Figure 1: Overall architecture of the SHIFT planner. Camera and LiDAR data are combined with an AI-based semantic map and robust VINS pose estimates to form a semantically enriched surface space. RFICP then generates an initial terrain-adaptive coverage trajectory, while IKD-SWOpt refines local paths in real time when encountering obstacles.
• Figure 2: An illustration of the local segment identification and refinement. For each waypoint on the initial path (red), a circular region (yellow) is defined. We then gather all identified non-compliment segments whose safety score exceeds the threshold into an adaptive sliding window for local optimization.
• Figure 3: Numerical simulation results: the SHIFT planner generates coverage trajectory that adapts to the environment and refines locally to avoid obstacles.
• Figure 4: A vacuum robot navigates a semantic-labeled indoor environment, slowing down in dirtier zones (red).
• Figure 5: Comparison of cleaning uniformity: SHIFT Planner achieves uniform cleaning in high-demand regions, while the other three methods show inadequate cleaning in heavily soiled areas.