Autonomous Navigation Stack for Industrial Cleaning Robot

This case study outlines the development of an indoor navigation stack for a commercial ride‑on floor‑scrubber platform. It focuses on the engineering approach, algorithms, and validation methodology. Specific vendor or product affiliations are intentionally omitted.

Goals

Reliable autonomous navigation in large indoor facilities (warehouses, retail, airports).
Repeatable routes (coverage paths) with human‑safe behavior and smooth motion.
Robustness to floor sheen, reflective obstacles, and mixed lighting.
Simple operator workflows: start/stop routes, pause/recover, docking.

Constraints

Indoor GNSS denial; reliance on onboard sensors and prior maps.
Variable traction and floor conditions (wet, polished, anti‑slip).
Safety requirements: speed limits, E‑stop integration, minimum stopping distance.
Compute budget aligned to embedded x86/ARM; deterministic control loops.

System Architecture

Middleware: ROS 2 (Humble/Foxy), composition nodes for modularity.
Navigation: Nav2 stack (global planner, local planner, recovery behaviors).
Perception: 2D lidar (primary), wheel odometry, IMU fusion; optional depth camera.
Mapping: Prior static map or online SLAM depending on site commissioning.
Control: Differential/drive‑by‑wire interface with velocity and steering commands.
Orchestration: Docker for deployment; health monitoring and logs.

Localization

Sensor fusion of wheel encoders + IMU via robot_localization (EKF/UKF).
Lidar scan‑matching (AMCL or NDT) against a 2D occupancy map for global pose.
Fail‑safes: pose continuity checks, covariance thresholds, relocalization triggers.

Mapping Options

Commissioned sites: Offline mapping pass with tuned lidar parameters; map cleaning and inflation for safety margins.
Dynamic sites: Online SLAM (e.g., Cartographer or Slam Toolbox) during teaching runs; export stabilized map for production.

Planning & Control

Global Planner: A* or Theta* on 2D occupancy with configurable cost weights.
Coverage Patterns: Lane‑based sweeping and serpentine paths for large halls.
Local Planner: DWB or TEB with obstacle inflation, velocity/acceleration limits, and footprint modeling.
Motion Quality: Jerk‑limited command smoothing; turn‑in‑place thresholds.

Obstacle Handling

Real‑time obstacle buffer around lidar detections with inflation for safety.
Dynamic obstacle anticipation via velocity obstacles (VO) heuristics at low speeds.
Recovery behaviors: clear costmap, rotate recovery, slow‑approach relocalization.

Safety Integration

Hardware E‑stop and safety PLC interfacing (read‑only status + command gating).
Speed caps near people‑dense zones; configurable geofences.
Minimum stopping distances based on friction estimates and payload.

Operator Workflow

Route selection from a curated list (pre‑taught or uploaded).
Start/pause/resume with clear audible/visual cues; graceful stop on interruptions.
Docking: approach path with slow final alignment; optional fiducials for precision.

Simulation & Testing

Digital twin in Gazebo/ignition with site‑like obstacles and friction maps.
Unit tests for planners; integration tests for localization loss/recovery.
KPIs: route completion rate, average lateral deviation, stop distance variance, relocalization count per hour.

Deployment

Containerized nodes with resource limits; log rotation; watchdogs for critical loops.
Configuration profiles per site (sensor offsets, inflation radius, speed caps).
Remote telemetry (optional): route summaries and fault codes (no PII).

Lessons Learned

Floor reflectivity impacts lidar; tune filtering and add redundancy where feasible.
Smooth, predictable motion increases human acceptance more than raw speed.
Clear operator UX (recover, resume, dock) reduces downtime significantly.

Note: This write‑up is an anonymized engineering overview. It does not imply affiliation with any specific manufacturer or model and excludes confidential details.