Milestone 1: Proposal & System Architecture

Mission Statement & Scope

Our robot is designed to autonomously explore a completely unknown indoor environment — such as a lab corridor or multi-room office space — without any prior map, human guidance, or pre-programmed paths. As it explores, it builds a live occupancy map and semantically classifies regions of the environment as clear, cluttered, or hazardous using its onboard RGB-D camera.

Target Environment: Static indoor environment — lab corridors, office rooms, and interconnected hallways. The environment contains furniture, walls, doorways, and open floor space. All obstacles are assumed static during a single exploration run.

Success State: The robot successfully explores at least 80% of the reachable floor area within a bounded environment, terminates exploration gracefully when no new frontiers exist, and produces a final labeled occupancy grid map distinguishing clear, cluttered, and hazardous zones — all without any human intervention after launch.

Technical Specifications

Robot Platform: TurtleBot4 Lite — iRobot Create3 base with Raspberry Pi 4 (2GB) compute board

Kinematic Model: Differential Drive — two independently driven wheels with a passive caster. The robot turns by varying the relative speed of the left and right wheels.

Perception Stack

High-Level System Architecture

The system follows the Perception → Estimation → Planning → Actuation pipeline. A reactive bypass allows the behavior coordinator to receive direct semantic feedback from the perception layer, enabling hazard-aware goal selection independently of map update speed.

🔍 Click to zoom

Module Declaration Table

Module Intent

Library Modules

SLAM Toolbox

We will use slam_toolbox in asynchronous mode (async_slam_toolbox_node) to generate a live 2D occupancy grid from the RPLIDAR scan data. Key configuration parameters: resolution (0.05m per cell), minimum_travel_distance and minimum_travel_heading (controlling map update frequency), and scan_buffer_maximum_scan_queue_size. SLAM Toolbox was chosen over RTAB-Map for its lower CPU overhead — critical given the Raspberry Pi 4's 2GB RAM constraints.

Robot Localization EKF

We will use the robot_localization package to implement an Extended Kalman Filter that fuses wheel encoder odometry (/odom at 20Hz) with the Create3's built-in IMU (/imu at 100Hz) and the OAK-D's secondary IMU (/oakd/imu/data). The EKF runs in 2D mode (estimating x, y, yaw only). The output /odometry/filtered topic provides SLAM Toolbox with a stable odom → base_link transform.

Nav2 Global Planner

Nav2 handles local path planning and obstacle avoidance once the Behavior Coordinator selects a frontier goal. We will tune inflation_radius (0.3m — one robot diameter) and configure the DWBLocalPlanner with max_vel_x: 0.2 m/s appropriate for indoor exploration. Nav2 integrates directly with the occupancy grid output from SLAM Toolbox and provides robust recovery behaviors when the robot gets stuck.

Diff-Drive Controller

The Create3 base directly consumes geometry_msgs/Twist messages on /cmd_vel, making the hardware interface transparent to ROS2. No custom firmware is required. Velocity and acceleration limits are configured through Nav2's controller parameters (max_vel_x: 0.22, max_vel_theta: 1.0, acc_lim_x: 0.5).

Custom Modules

Frontier Explorer Node

This node implements a frontier-based exploration algorithm from scratch, based on the foundational work of Yamauchi (1997) extended with information-theoretic goal scoring. The node subscribes to /map (the live nav_msgs/OccupancyGrid) and identifies frontier cells — free cells (value 0) adjacent to at least one unknown cell (value -1). It uses connected-component labeling via scipy.ndimage to cluster frontier cells into contiguous regions.

Each frontier region is scored using an information-gain function:

The node publishes a ranked geometry_msgs/PoseArray to /frontier_goals and a visualization_msgs/MarkerArray to /frontier_markers for live RViz2 inspection.

Semantic Hazard Classifier

This node processes the OAK-D camera feed to classify regions into three semantic categories: CLEAR (open passable space), CLUTTERED (moderate obstacle density), and HAZARD (dense obstacles, transparent surfaces, or potential drop-offs). Each frame is divided into a spatial grid of patches. For each patch: edge density via Canny detection on the RGB image, mean depth and depth variance from the stereo image, and the ratio of NaN depth values. These features feed a rule-based classifier tuned on labeled lab images. Output is published as a coarse nav_msgs/OccupancyGrid to /semantic_map.

Behavior Coordinator

This node implements the state machine governing all exploration decisions. It defines five states: IDLE → SELECTING → NAVIGATING → ARRIVED → DONE. It subscribes to /frontier_goals, /semantic_map, and /battery_state, and uses the Nav2 NavigateToPose action client to dispatch goals. The coordinator filters frontier candidates against the current semantic map — skipping any frontier overlapping a HAZARD zone — and maintains a visited-frontiers set to prevent re-exploring.

Safety & Operational Protocol

Deadman Switch: If no /frontier_goals update is received for more than 5 seconds, the coordinator publishes zero-velocity to /cmd_vel and transitions to IDLE.

Git Infrastructure

Team Repository: github.com/PriColon/mobile-robotics-frontier-exploration ↗

All three teammates added as members with write access. Each team member's personal project site includes the shared repository as a Git submodule at _projects/tb4-frontier-exploration.