💻 computer science

Semantic-Aware Autonomous Exploration for UAVs in Unknown Indoor Environments

This paper proposes a semantic-aware autonomous exploration framework for UAVs that enhances a roadmap-based strategy with a semantic reward function to prioritize meaningful objects, thereby achieving higher coverage rates (90–94%) with reduced time and travel distance compared to conventional geometry-based methods.

Original authors: Duc-Thien Nguyen, Ngoc Minh Do, Xiem HoangVan, Thanh Nguyen Canh

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Duc-Thien Nguyen, Ngoc Minh Do, Xiem HoangVan, Thanh Nguyen Canh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a drone (a flying robot) that needs to map out a house it has never seen before. Usually, these drones act like a person walking through a dark room with a flashlight, checking every single inch of empty air just to make sure it's empty. They treat a blank wall, a hallway, and a pile of furniture exactly the same way: "I haven't seen this yet, so I need to look at it."

This paper proposes a smarter way for the drone to think. Instead of just looking at where things are (geometry), the drone learns to recognize what things are (semantics).

Here is how the system works, broken down into simple concepts:

1. The "Smart Tour Guide" vs. The "Random Walker"

The Old Way: Imagine a tourist in a new city who decides to visit every single street, alley, and empty lot with the same level of excitement. They might spend an hour staring at a blank wall just because they haven't seen it yet, while missing the famous museum down the street. This is how most current drones work; they treat all unknown space as equally important.

The New Way: This paper gives the drone a "Smart Tour Guide." This guide knows that some places are more interesting than others.

The "Semantic" Layer: The drone uses a camera to recognize objects like doors, corridors, or specific items.
The "Reward" System: Think of this like a video game. If the drone flies near a door or a hallway, it gets a "bonus point" (a semantic reward). If it flies near a blank wall, it gets zero points.
The Result: The drone naturally wants to go where the "bonus points" are. It prioritizes exploring doorways and rooms over empty, boring spaces, making the job much faster.

2. The "Roadmap" (The Mental Map)

To navigate, the drone doesn't just fly randomly; it builds a mental "roadmap" (called a Probabilistic Roadmap or PRM).

How it works: Imagine the drone is drawing a web of possible paths in the air. It constantly adds new dots (possible places to fly) and connects them with lines (safe paths) as it learns more about the room.
The Upgrade: In the past, this map only knew about walls and obstacles. Now, the map also knows, "Hey, there's a door over there!" and highlights that part of the web as a high-priority destination.

3. The Balancing Act

The drone has to be careful. It can't just fly straight to the most interesting object and crash into a wall. It uses a formula to balance three things:

How much new information will I get? (Geometric gain)
How interesting is this spot? (Semantic reward)
How far do I have to fly? (Travel cost)

It picks the spot that gives the best "bang for the buck"—high interest, good new info, but not too far away.

4. The Results: Faster and Smarter

The authors tested this in three different simulated apartment sizes (small, medium, and large).

The Outcome: The "Smart Tour Guide" drone finished mapping the rooms faster and flew a shorter distance than the "Random Walker" drones.
The Numbers: It managed to map between 90% and 94% of the rooms. In the biggest, most complex apartment, it was significantly faster than the other methods, cutting down the time needed to explore by about 17% to 34% compared to standard methods.
Why it won: It didn't waste time flying in circles around empty spaces. It went straight for the "interesting" parts of the house first.

Summary

In short, this paper teaches a drone to stop being a robot that just scans empty air and start acting like a curious human who knows that doors, hallways, and objects are more important to find than empty space. By giving the drone a sense of "what matters," it can explore unknown buildings much more efficiently.

Technical Summary: Semantic-Aware Autonomous Exploration for UAVs in Unknown Indoor Environments

Problem Statement
Autonomous exploration in unknown environments requires Unmanned Aerial Vehicles (UAVs) to generate informative trajectories while constructing accurate maps. While existing methods, particularly sampling-based planners like the Dynamic Exploration Planner (DEP), utilize geometric information (e.g., occupancy and voxel maps) to evaluate candidate viewpoints, they often lack semantic awareness. Consequently, these methods treat all unexplored regions as equally important, regardless of their relevance to specific tasks. This limitation leads to suboptimal exploration efficiency and redundant motion, as the UAV may prioritize ordinary free space over semantically significant areas such as doorways, corridors, or object-rich regions.

Methodology
The paper proposes a semantic-aware exploration framework that augments a roadmap-based strategy with a semantic layer. The system operates in a closed loop using three primary modules:

Map Module: Constructs and maintains a 3D occupancy representation using a log-odds formulation based on RGB-D sensor data. It simultaneously generates an Euclidean Signed Distance Field (ESDF) to ensure safe trajectory generation.
Semantic Module: This module introduces task-oriented exploration by detecting semantic objects and assigning them importance weights. For each detected object, a semantic region is defined by its centroid, label, importance weight, and influence radius. A semantic reward function is calculated for any roadmap node based on its proximity to these regions; the reward decreases linearly with distance and vanishes outside the influence radius.
Roadmap Management Module: Based on the DEP, this module incrementally constructs a Probabilistic Roadmap (PRM). It continuously updates the graph by inserting the current UAV position, sampling new candidate nodes in unexplored space, and connecting them via collision-free edges.

Viewpoint Evaluation and Planning
The core innovation lies in the utility function used to select the next viewpoint. The total utility $U(p)$ combines the geometric information gain $G(p)$ with the semantic reward $R_{sem}(p)$ :
$U(p) = G(p) + R_{sem}(p)$
The geometric gain is computed based on the number of observable normal, frontier, and surface voxels. The final score for a candidate node is this combined utility normalized by the estimated travel cost (path length and yaw rotation). This ensures the UAV prioritizes regions that are both geometrically informative and semantically meaningful, while favoring nearby targets to minimize travel time.

Key Contributions
The authors summarize their contributions as follows:

Framework Integration: A semantic-aware exploration framework that integrates semantic information into a DEP-based architecture.
Reward Mechanism: A semantic reward mechanism that prioritizes exploration targets according to their contextual significance (e.g., doors, intersections, objects).
Experimental Validation: Evaluation in multiple simulated indoor environments demonstrating improved exploration efficiency compared to conventional geometry-based methods.

Experimental Results
The framework was implemented in ROS Noetic and Gazebo using a quadrotor UAV equipped with an RGB-D camera. Experiments were conducted in three environments of varying complexity (50m², 100m², and 200m²). The proposed method was compared against existing approaches: Next-Best View (NBV), Autonomous Exploration Planner (AEP), and the baseline DEP.

Coverage: The proposed approach achieved map coverage rates between 90% and 94% across all environments.
Efficiency: Compared to conventional methods, the proposed framework significantly reduced exploration time and travel distance.
- In the largest environment (200m²), the method achieved 90.42% coverage, outperforming DEP, NBV, and AEP by 7.15, 39.31, and 10.67 percentage points, respectively.
- Exploration time was reduced by 16.7% compared to DEP and 33.9% compared to AEP in the complex environment.
Sampling Density: The study identified that an appropriate sampling density (e.g., 30–45 nodes depending on environment size) is critical; excessive sampling increases computational overhead without improving coverage, while insufficient sampling degrades performance.

Significance and Claims
The paper claims that by jointly considering geometric information gain and semantic importance, the proposed framework enables UAVs to explore unknown indoor environments more effectively than methods relying solely on geometry. The integration of semantic rewards allows the UAV to focus on areas with higher information value, thereby reducing redundant motion. The authors note that future work will focus on integrating real-time semantic perception using deep learning models, validating the framework on real UAV platforms, and investigating adaptive semantic weighting and multi-UAV cooperative exploration.