← Back to Home

Autonomous Mapping — CALL-M Mobile Manipulator

ROS 2SLAMNav2Frontier ExplorationLiDARRGB-DDynamic Environments
GitHub View Documentation

Overview

At the CNRS-AIST Joint Robotics Laboratory in Tsukuba, I developed a modular ROS 2 architecture for autonomous mapping with CALL-M, an omnidirectional mobile manipulator. The goal was to move beyond running SLAM under manual driving and build a system that initializes a distributed robotics stack, fuses heterogeneous range sensors, builds an occupancy map, decides where to explore next, supervises navigation goals, recovers from failures, handles moving obstacles, tracks mission progress, and terminates the mission in a controlled way.

The project produced three contributions: a supervised ROS 2 software architecture for complete autonomous-mapping missions, a reusable exploration-management framework for frontier selection and goal supervision, and a static–dynamic obstacle-filtering framework for separating persistent structure from moving observations.

System Architecture

The mapping pipeline connects five components: multimodal scan fusion, static–dynamic obstacle filtering, SLAM Toolbox, frontier-based exploration, and Nav2. CALL-M combines two RPLIDAR A1M8 scanners, two ZED Mini stereo cameras, an Intel RealSense D435i, an omnidirectional TriOrb base, and a Universal Robots manipulator, running as a distributed ROS 2 graph across an NVIDIA Jetson Orin Nano, an Intel NUC, and an operator computer.

Contribution 1 — Supervised Mapping Architecture

A program manager coordinates the complete mission lifecycle: launching the distributed stack in dependency order, waiting for fused sensor data and a ready occupancy map, gating autonomous exploration behind explicit confirmation, coordinating execution across all three computers, preserving the final map, and shutting the system down in an orderly sequence. The contribution here is not a new SLAM or navigation algorithm — it is the software architecture that makes established components operate together as one reliable autonomous-mapping mission.

Contribution 2 — Exploration-Management Framework

I built a reusable exploration-management framework that extends frontier exploration past “go to the nearest unknown area.” It extracts and clusters frontier regions, validates and scores candidates, dispatches and supervises Nav2 goals, maintains spatial memory of failed targets, and decides when the mission is complete — requiring only an occupancy map, TF-based pose, and a Nav2-compatible interface, so it transfers to other ROS 2 mobile robots through configuration alone.

Contribution 3 — Static–Dynamic Obstacle Filtering

I also developed a ROS 2 preprocessing framework that separates persistent environmental structure from moving observations before they reach SLAM, producing a static scan for map construction and a dynamic scan that stays available to navigation for collision avoidance. The goal isn’t to hide people from the robot — it’s to keep people from becoming walls in the map. This framework is being developed as an independent research contribution.

Results

Simulation:

  • 10/10 autonomous mission completions with the supervised architecture, vs. 7/10 for a nearest-frontier baseline.
  • 95.3 ± 2.4% coverage of reachable space.
  • Average navigation goals reduced from 161.9 to 56.1.
  • Average progress per goal increased from 1.21 m to 3.86 m.

Real robot:

  • 5/5 autonomous mapping sessions reached completion.
  • A measured 4.90 m room dimension was reconstructed as 4.85 ± 0.05 m.
  • A measured 6.80 m dimension was reconstructed as 7.13 ± 0.07 m.

Research Output