Milestone 2: Mid-Point Technical Proof

Overview

Milestone 2 documents our transition from system design (M1) to working implementation. This milestone presents concrete evidence of three custom ROS2 nodes running on the TurtleBot4 Lite hardware platform, a complete single-command launch system, and a recorded ROS bag demonstrating live frontier exploration with semantic hazard classification.

Our system enables a TurtleBot4 Lite to autonomously explore unknown indoor environments, detect and classify environmental hazards using camera, depth, LiDAR, and onboard sensors, and navigate safely by avoiding dangerous frontier zones — all without human intervention.

System Architecture

The full exploration stack connects four primary subsystems: SLAM Toolbox builds the occupancy map, the frontier explorer identifies exploration targets, the semantic hazard classifier labels dangerous zones, and the behavior coordinator orchestrates safe navigation through Nav2.

Frontier Explorer Node

The frontier explorer node (frontier_explorer_node.py), developed by Princess Colon, implements the core exploration strategy. It continuously processes the SLAM occupancy map to identify unexplored regions and publishes ranked navigation targets to the behavior coordinator.

Algorithm

The node subscribes to /map (OccupancyGrid from slam_toolbox) and /odometry/filtered (EKF-filtered pose). Every time a new map arrives, the following pipeline executes:

The scoring formula balances three competing objectives: size rewards larger unknown regions, unknown_near rewards frontiers surrounded by more undiscovered space (maximizing information gain), and dist penalizes faraway frontiers to minimize unnecessary travel. The parameters α=1.0 and β=2.0 weight unknown area discovery twice as heavily as cluster size, reflecting the goal of maximizing map coverage.

Evidence

Semantic Hazard Classifier Node

The semantic hazard classifier (semantic_hazard_classifier_node.py) is the core contribution of this milestone from Manjunath Kondamu. It runs at 10 Hz and processes data from five independent sensor layers simultaneously, fusing their outputs into a single hazard classification that is published to the robot's semantic map.

The term "semantic" refers to attaching meaningful labels to spatial regions rather than simply marking them as occupied or free. Where a standard occupancy grid treats a wall and a fire identically, our semantic map distinguishes between CLEAR, DEBRIS, SMOKE, WATER, DARK, NARROW, POTHOLE, GLASS, HAZMAT, FIRE, and CLIFF zones — enabling the behavior coordinator to make informed navigation decisions.

5-Layer Detection Architecture

Hazard Priority Fusion

All five layers run in parallel every 100 ms. Their detections are collected into a list and passed to fuse_hazards(), which selects the single highest-priority hazard using a predefined severity scale:

HAZARD_PRIORITY = {
    'CLIFF':    100,   # immediate physical danger
    'FIRE':     100,   # immediate physical danger
    'POTHOLE':  90,    # structural floor hazard
    'GLASS':    85,    # invisible obstacle
    'HAZMAT':   85,    # chemical/biological risk
    'WATER':    80,    # slippery surface / flooding
    'DEAD_END': 80,    # navigation trap
    'NARROW':   70,    # passage risk
    'SMOKE':    60,    # visibility / air quality
    'DARK':     60,    # unknown danger
    'DEBRIS':   40,    # passable but cautious
    'CLEAR':    0,
}

Critical hazards (CLIFF, FIRE, POTHOLE, HAZMAT) trigger an immediate E-stop by publishing True to /exploration/stop, halting the behavior coordinator and zeroing the velocity command.

DeepHAZMAT Integration

Layer 5 integrates the DeepHAZMAT framework (Najafi et al., 2021), which uses a YOLOv3-tiny architecture trained on the HAZMAT-13 dataset containing 52,845 images across 13 hazardous material sign categories. The model was selected specifically for its design under restricted computational resources — matching the constraints of the Raspberry Pi 4 (2 GB RAM) aboard the TurtleBot4 Lite.

DeepHAZMAT Detection — Evidence

The following screenshots demonstrate DeepHAZMAT YOLOv3-tiny running on sample HAZMAT images, confirming successful detection and bounding box generation prior to robot camera integration.

Outputs

The classifier publishes to three topics simultaneously:

• /semantic_map — OccupancyGrid where cell values encode hazard severity (0=CLEAR, 30=DEBRIS, 85=HAZMAT, 100=FIRE/CLIFF)
• /hazard_alert — String with current highest-priority hazard name
• /hazard_markers — MarkerArray of colored 3D cubes for RViz2 visualization

Behavior Coordinator Node

The behavior coordinator (behavior_coordinator_node.py) is the decision-making brain of the exploration system. It integrates the outputs of both the frontier explorer and the semantic hazard classifier to select safe navigation goals and drive the robot to full map coverage.

5-State Machine

When no valid frontiers remain (all visited or all hazardous), the node transitions to DONE, publishes a zero velocity command, and emits EXPLORATION_COMPLETE on /exploration/status.

Safety Systems

Launch System

The entire exploration stack launches with a single command via exploration.launch.py. The launch file starts all subsystems in the correct order with appropriate delays:

ros2 launch frontier_exploration_mapping exploration.launch.py

The use_sim_time:=true argument switches all nodes to Gazebo simulation time for testing, while the default false uses hardware time for deployment on the physical TurtleBot4 Lite.

ROS Bag Recording

A complete ROS bag was recorded during the frontier exploration demo run. The bag captures all sensor and algorithmic topics required for post-analysis and replay.

To replay the bag:

ros2 bag play ros2_ws/rosbags/rosbag_full_frontier_functional

Mathematical Foundations

The mathematical derivations supporting all algorithmic decisions in this system were compiled by Rohit Mane and are provided in full as LaTeX-formatted documents. Key derivations include the frontier scoring function, EKF state estimation equations, differential drive kinematics, occupancy grid Bayesian update rules, and the HSV threshold derivations for the semantic classifier.

Team Contributions

The following table documents the specific technical contributions of each team member for Milestone 2.

Next Steps — Milestone 3