Autonomous Drone-Based Warehouse Inspection

3D Grid Mapping and A* Path Planning for Autonomous Rack Inspection in Simulated Warehouse Environments

Abstract

This project presents the design and implementation of an autonomous drone navigation system for warehouse inspection tasks. The system constructs a 3D occupancy grid of a simulated warehouse environment and uses a custom 3D A* path planning algorithm to navigate a drone efficiently between rack locations for item inspection. The approach integrates stereo camera-based perception, ROS/ROS2 middleware, and a Euclidean distance heuristic to compute shortest traversal paths through three-dimensional space, enabling safe and scalable autonomous inspection without human intervention.

1. Introduction

Modern warehouses require frequent inspection of goods stored on multi-level racking systems — a task that is time-consuming and physically demanding when performed manually. Autonomous aerial robots (drones) offer a compelling alternative, capable of navigating three-dimensional spaces and inspecting rack faces at multiple heights without scaffolding or human risk. This project investigates the core algorithmic challenges of deploying such a system: building a 3D spatial representation of the environment and computing collision-free, shortest-distance inspection paths using a 3D A* search algorithm implemented in C++ within the ROS ecosystem.

2. System Overview

3D Grid Mapping: The warehouse environment is discretized into a 3D voxel grid. Each voxel stores occupancy probability derived from stereo camera depth estimates, representing free space, obstacles (racks, pillars, floor), and unknown regions.
3D A* Path Planner: A custom A* implementation searches the 3D voxel grid from the drone's current position to each inspection waypoint, using 3D Euclidean distance as the heuristic to minimize total traversal distance.
ROS/ROS2 Integration: The mapping, planning, and control nodes communicate over ROS topics and services, enabling modular development and real-time replanning.
Stereo Camera Perception: A stereo camera onboard the ModelAI Voxel 2 provides disparity-based depth data for both grid updates and obstacle detection during flight.
Simulation Environment: The full pipeline is validated in a Gazebo-based simulated warehouse with configurable rack layouts before physical deployment.

Hardware

ModelAI Voxel 2 Drone
Stereo Camera
Onboard compute unit

Software & Tools

ROS / ROS2
C++ / Python
Gazebo Simulation
Git, Docker

3. Methodology

3.1 3D Voxel Grid Construction

The warehouse environment is represented as a 3D occupancy grid where each cell (voxel) is a fixed-size cube in metric space. As the drone flies, stereo camera depth frames are back-projected into 3D point clouds and used to update voxel occupancy probabilities using a log-odds update rule. Free voxels form the navigable space; occupied and unknown voxels are treated as obstacles for planning purposes.

Voxel resolution is configurable to trade off map fidelity against memory footprint.
The grid is bounded to the warehouse extents and initialized as unknown prior to exploration.
Stereo disparity is converted to depth using the camera's calibrated baseline and focal length.

3.2 3D A* Path Planning Algorithm

The core research contribution is a 3D A* search operating directly on the voxel grid. The algorithm expands nodes in 26-connected neighborhoods (faces, edges, and corners of each voxel), evaluating movement cost and a Euclidean heuristic to find the shortest collision-free path from the drone's current voxel to the target inspection voxel.

3D Euclidean Heuristic: h(n) = √( Δx² + Δy² + Δz² ) — admissible and consistent, guaranteeing optimal paths.
26-connectivity: Diagonal and vertical moves are permitted, enabling the planner to exploit free airspace and minimize detours.
Shortest Traversal Distance: The planner minimizes total flight distance, reducing inspection time and battery consumption.
Paths are smoothed post-planning to remove unnecessary heading changes before being sent to the drone's flight controller as waypoint sequences.

Research Focus: Development and integration of the 3D A* path planning algorithm — including 3D grid representation, 3D Euclidean distance calculation, and shortest traversal distance optimization across the full voxel graph.

3.3 Inspection Waypoint Sequencing

Rack inspection points are pre-defined as 3D coordinates corresponding to the face of each shelf segment. The system sequences these waypoints to minimize total flight path length — analogous to a constrained coverage problem — before invoking the A* planner between consecutive inspection goals.

Inspection targets are loaded from a YAML configuration file defining rack positions and heights.
A greedy nearest-neighbor ordering is applied to sequence waypoints by proximity.
The 3D A* planner computes the collision-free path between each consecutive waypoint pair.
The drone executes each path segment, pausing at each waypoint to capture and log inspection imagery.

3.4 System Pipeline

01 Map Initialization: Warehouse bounds and rack positions loaded; 3D voxel grid initialized to unknown state.
02 Exploration Pass: Drone performs a pre-programmed sweep to populate the voxel grid with stereo camera depth data.
03 Waypoint Planning: Inspection waypoints are sequenced; 3D A* computes shortest collision-free paths between each pair.
04 Autonomous Flight: Drone executes waypoint paths, updating the occupancy grid continuously and replanning if new obstacles are detected.
05 Inspection & Logging: At each rack face, stereo imagery is captured and logged for downstream inventory analysis.

3.5 Simulation Environment

The full system is tested in a Gazebo simulation of a warehouse with configurable rack layouts, aisle widths, and ceiling heights. The ModelAI Voxel 2 drone model is simulated with realistic physics and stereo camera plugin output, enabling direct transfer of the ROS nodes to the physical platform without code modification.

Gazebo warehouse simulation with drone and racks — Fig. 1. Simulated warehouse environment in Gazebo showing drone, rack layout, and planned inspection path

3D occupancy voxel grid visualization — Fig. 2. 3D voxel occupancy grid constructed from stereo camera depth data — free (blue), occupied (red), unknown (grey)

3D A* path visualized through voxel grid — Fig. 3. 3D A* shortest path (green) computed through the voxel grid between two rack inspection points

4. Results

The system successfully demonstrated end-to-end autonomous warehouse inspection in simulation. The 3D A* planner consistently found shortest collision-free paths through the voxel grid, with the Euclidean heuristic providing near-optimal path quality across varying rack configurations. Replanning on newly observed obstacles completed within real-time constraints. Inspection coverage of all defined rack faces was achieved in each trial without any collisions.

3D A* path search executed within planning time budgets suitable for online replanning.
Euclidean heuristic reduced node expansions significantly compared to uninformed search.
Full rack coverage achieved across all simulated warehouse layouts tested.
Stereo-based voxel updates maintained consistent map fidelity throughout flight.

Autonomous Drone Inspection Demo — 3D A* navigation through simulated warehouse (in progress)

5. Conclusion

The 3D A* path planning algorithm and all associated concepts — including voxel grid construction, Euclidean distance heuristics, and stereo camera-based occupancy mapping — have been successfully tested and validated within the Gazebo simulation environment. The results confirm the correctness and efficiency of the approach in representative warehouse conditions. The next phase of this project is to replicate and deploy the same system in a real-world warehouse setup, using the ModelAI Voxel 2 drone equipped with a stereo camera, operating over physical rack structures. This transition from simulation to reality will validate the full pipeline under real sensor noise, lighting conditions, and dynamic environments.