SAFARI: Semantic- And Frontier-driven Autonomous Risk Inspection

Ugur Kafkasyali; Mihnea Toader; Julia Richter; Turcan Tuna; Cesar Cadena; Marco Hutter

cs.ROOct 30, 2026

SAFARI: Semantic- And Frontier-driven Autonomous Risk Inspection

Authors: Ugur Kafkasyali, Mihnea Toader, Julia Richter, Turcan Tuna, Cesar Cadena, Marco Hutter

Organizations: Robotic Systems Lab, Swiss Federal Institute of Technology (ETH Zürich), 8092 Zürich, Switzerland

Abstract

Robotic exploration can reduce human exposure to hazardous environments by autonomously searching unknown areas for objects of interest while constructing a complete, risk-aware map. However, existing autonomous exploration approaches typically prioritize geometric coverage or information gain, without explicitly coupling semantic object detection and the risk associated with operating near potentially hazardous objects. Moreover, many systems have been demonstrated primarily in simulation or controlled environments, leaving limited validation of risk-aware semantic exploration in complex, unstructured field settings. To address this gap, we propose SAFARI, a framework that treats object detection as a native channel of a unified probabilistic map, thereby jointly informing exploration and planning with both geometric and semantic evidence. We evaluated SAFARI in 11 missions across three field sites on a quadruped platform. The framework achieved an average coverage of 98.7% of the accessible mission area, without operator intervention for 97.6% of mission time, and detected objects with characteristic dimensions as small as 15cm. Lastly, through ablation studies and a detailed analysis of system behavior, we show that object-size awareness and safety-aware path planning are crucial for effective hazardous-object search.

Figures & tables

Figure 1: Illustration of the mission setup. Given a bounded region, the robot autonomously explores, while simultaneously detecting and avoiding collision with objects of potential interest. (A) Third-person view of an ANYbotics ANYmal D approaching and correctly detecting two OPIs: a trash can and a littered bottle. (B) First-person view of the robot’s perspective, overlaid with segmentation masks and detection output. (C) Accumulated risk layer overlaid on satellite imagery of the explored area. The robot’s trajectory is depicted in orange.

Figure 2: System overview diagram of the SAFARI framework. It depicts the flow of information through the detection, mapping and navigation modules, as well as the operating concept for each. It combines real data (sensor data, detector output, map layers) and illustrations (robot, viewpoints, paths) for a more comprehensive depiction.

Figure 3: Map and traversability layers overlaid on the mission footage. Panels (A) and (B) show the geometric layer, while panels (C) and (D) show the hazard layer. (A) The terrain-grade estimate $\theta_{i}^{\mathrm{grade}}$ , representing the large-scale inclination of the estimated support surface. (B) The residual-slope estimate $\theta_{i}^{\mathrm{res}}$ , highlighting local height discontinuities and protrusions after removing the broader terrain grade. (C) The hazard lambda field intensity $\lambda^{\mathrm{haz}}_{i,k}$ , representing the spatial intensity derived from the accumulated positive and negative risk evidence for the active size hypothesis. (D) The corresponding Beta variance $V_{i,k}$ , used as an uncertainty measure for the accumulated hazard-evidence rate.

Symbol	Description	Value
Geometric layer
$\epsilon_{\mathrm{geom}}$	Error-region footprint area	$0.10\times 0.10\,\mathrm{m}^{2}$
$\alpha_{0}^{\mathrm{geom}}$	Hit prior count	$0.5$
$\beta_{0}^{\mathrm{geom}}$	Miss prior count	$25.0$
Hazard layer
$A_{\mathrm{haz}}$	Hazard error-region area	$0.25\times 0.25\,\mathrm{m}^{2}$

Table 1: Mapping parameters used.

Figure 4: Figure demonstrating the OPI evidence accumulation of the lambda map. (A) Updates to the OPI layer of the lambda map at different time steps of a mission. At $t_{2}$ , a false detection is simulated to show how the mapper recovers from it. (B) Changes in the $\lambda_{i}$ values of two selected map cells. (C) The expected OPI size mechanism. Small and medium OPIs are not detected beyond the distances $s_{0}$ and $s_{1}$ , respectively.

Figure 5: A summary of the steps of the path planning process. (A) Formation of frontier points. (B) Construction of a rough path towards the targeted frontier point. (C) Candidate viewpoints, their respective paths, and the chosen path. (D) Map and planning visuals overlaid on aerial imagery of a mission area.

Figure 6: The ANYmal D quadruped deployed for the missions. It is equipped with a Stereolabs ZED X depth camera, an NVIDIA Jetson AGX Orin, and a GNSS unit.

Figure 7: Overview of per-site tasks and primary mission outputs. The leftmost column overlays the path of the robot in the baseline missions, the location of OPIs, and the coverage over time curve. The middle column displays the hazard layer, traversability layer, and accumulated 3-D geometry for each mission. The rightmost column correlates the numbered OPI labels with the actual object, highlighted in yellow. (A) Roadside cleanup; (B) storage tank inspection; (C) disaster response.

Mission	System metrics						Coverage			Detections		Interventions
Mission	Speed	Stand	Path	Dur.	Area	Access.	MCPS	50/75/90%	Final	Detections	False Det.	Num.	Duration	No-Int.
	[ $\mathrm{m}\text{/}\mathrm{s}$ ]	[ $\mathrm{s}$ ]	[ $\mathrm{m}$ ]	[ $\mathrm{s}$ ]	[ ${\mathrm{m}}^{2}$ ]	[ ${\mathrm{m}}^{2}$ ]	[ ${\mathrm{m}}^{2}\text{/}\mathrm{s}$ ]	[ $\mathrm{s}$ ]	[%]	[#]	[#]	[#]	[ $\mathrm{s}$ ]	[%]
M01	0.39	198	431.5	1105	319	282	0.25	28/78/180	99.98	8/8	0	3	8.5	98
M02	0.44	81	216.2	495	320	263	0.53	38/167/229	99.44	6/8	0	6	13.7	83
M03	0.42	89	312.7	744	317	275	0.36	41/107/235	99.92	8/8	0	0	–	100
M04	0.41	54	191.8	471	318	272	0.57	54/134/154	99.60	6/8	0	0	–	100

Table 2: Per-mission metrics: system, coverage, detection, and intervention statistics. Speed is path length divided by total recorded duration, including standing time. Standing uses a

0.05\text{\,}\mathrm{m}\text{/}\mathrm{s}

threshold; Access. is the observable subset of the mission area; intervention duration is averaged per takeover.

Ablation	Mission	Location	$\alpha$	Obj. radius [m]	PTU
Baseline (reference config.)	M03	road	0.0	0.2	off
	M07	tanks	0.0	0.2	off
	M10	building	0.0	0.2	off
Explor./exploit. dial (varies $\alpha$ )	M01 †	road	0.3	0.2	on
	M05	road	0.5	0.2	off
	M04	road	1.0	0.2	off

Table 3: Index of all performed ablations. Parameters that differ from the baseline are shown in bold. † M01 varies two factors simultaneously.

Figure 8: Accumulated risk layer of the M03 mission, shown without clipping to the mission polygon (blue). The numbered OPI labels correspond to the visual references from Figure 7 . (A) OPI next to an untraversable granite block; (B) cluster of three OPIs.

Panel of Figure 9 — Figure 9: Effect of the exploration-exploitation dial $\alpha$ on coverage rate. $\alpha=0$ selects frontiers purely by exploration demand; $\alpha=1$ penalizes travel distance and therefore prefers nearby, already-accessible frontiers.

Figure 10: Coverage curves with the PTU off and on (road: M03 off vs. M02 on; building: M10 off vs. M11 on). The runs show no consistent coverage-rate advantage from enabling the PTU.

Panel of Figure 11 — Figure 11: Effect of the active hazard-size hypothesis $R_{k^{*}}$ on search behavior, for two road missions with otherwise identical parameter settings. The size hypothesis sets the range-dependent weighting of negative evidence. Panel (a) shows the corresponding coverage curves.

Figure 12: Per-frame detection rate against range to the object, pooled by object class over the missions at the road and tank sites.

Autonomous Robot Exploration

See all 38 papers

Semantic-Aware Autonomous Exploration for UAVs in Unknown Indoor Environments
Jun 21, 2026Duc-Thien Nguyen, Ngoc Minh Do, Xiem HoangVan +1Robot Navigation Informative Path Planning
University of Engineering and Technology, Vietnam National University, 10000, Hanoi, Vietnam. · School of Information Science, Japan Advanced Institute of Science and Technology, Nomi, 923-1211, Ishikawa, Japan.
SEAMLiS: Visibility-Aware Safety for Perception-Limited Multi-Robot Exploration
Jul 10, 2026Taekyung Kim, Rahul H Kumar, Aswin D. Menon +2Robotic Control Multi-Robot Exploration
Department of Robotics, University of Michigan, Ann Arbor, MI, 48109, USA · Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI, 48109, USA · Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI, 48109, USA
RA-Nav: A Risk-Aware Navigation System Based on Semantic Segmentation for Aerial Robots in Unpredictable Environments
Feb 19, 2026Ziyi Zong, Xin Dong, Jinwu Xiang +2Safe Robot Navigation Safe Motion Planning

Jun 21, 2026J/K move · Enter open · S save

Semantic-Aware Autonomous Exploration for UAVs in Unknown Indoor Environments

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

University of Engineering and Technology, Vietnam National University, 10000, Hanoi, Vietnam. · School of Information Science, Japan Advanced Institute of Science and Technology, Nomi, 923-1211, Ishikawa, Japan.

Robot Navigation Informative Path Planning UAV Navigation Sampling-Based Motion Planning

Open paper AskPDF arXiv

Autonomous exploration in unknown environments requires unmanned aerial vehicles (UAVs) to efficiently generate informative trajectories while simultaneously constructing accurate maps. Although many existing exploration methods rely on geometric information, they often lack semantic awareness, resulting in suboptimal exploration efficiency and limited environmental understanding. To address this limitation, this paper proposes a semantic-aware exploration framework that adds semantic information to a roadmap-based exploration strategy. The proposed method builds on the Dynamic Exploration Planner (DEP), which incrementally constructs a Probabilistic Roadmap (PRM), and augments this roadmap with a semantic layer. A semantic reward function is introduced to prioritize regions containing meaningful objects and structures, enabling the UAV to focus on areas with higher information value. Furthermore, the roadmap is continuously updated to support efficient frontier selection and path planning during exploration. The proposed framework is implemented in ROS Noetic and Gazebo using an RGB-D sensor for simultaneous acquisition of geometric and semantic information. Experimental results in multiple simulated environments demonstrate that the proposed approach achieves exploration coverage rates between 90% and 94% while reducing exploration time and travel distance compared with conventional geometry-based exploration methods.

Abstract

Autonomous Robot Exploration

Explore similar work

Semantic-Aware Autonomous Exploration for UAVs in Unknown Indoor Environments

SEAMLiS: Visibility-Aware Safety for Perception-Limited Multi-Robot Exploration

RA-Nav: A Risk-Aware Navigation System Based on Semantic Segmentation for Aerial Robots in Unpredictable Environments