Learning Semantic Priorities for Autonomous Target Search
Authors: Max Lodel, Nils Wilde, Robert Babuška, Javier Alonso-Mora
Organizations: Department of Cognitive Robotics (CoR), Delft University of Technology, The Netherlands · CIIRC, Czech Technical University in Prague, Czech Republic
The use of semantic features can improve the efficiency of target search in unknown environments for robotic search and rescue missions. Current target search methods rely on training with large datasets of similar domains, which limits the adaptability to diverse environments. However, human experts possess high-level knowledge about semantic relationships necessary to effectively guide a robot during target search missions in diverse and previously unseen environments. In this paper, we propose a target search method that leverages expert input to train a model of semantic priorities. By employing the learned priorities in a frontier exploration planner using combinatorial optimization, our approach achieves efficient target search driven by semantic features while ensuring robustness and complete coverage. The proposed semantic priority model is trained with several synthetic datasets of simulated expert guidance for target search. Simulation tests in previously unseen environments show that our method consistently achieves faster target recovery than a coverage-driven exploration planner.
Figures & tables
Fig. 1 : Conceptual overview of the proposed framework. During data collection, an expert generates interventions into the planner’s goal output, prioritizing certain semantically relevant objects (depicted by stars). These are used to train a semantic model, which outputs priorities for each exploration frontier that, in turn, guide the exploration planner. The exploration planner outputs the next frontier viewpoint to navigate to.
Fig. 2 : Top-down views of an example scenario of the first task setup comparing coverage-driven exploration with our learned semantic priority model. Frontier nodes of the topological graph are colored black, and others are gray. Blue edges visualize the path taken by the robot; the larger blue circle is the robot’s position at target discovery time, and the smaller blue circle is the initial position. The red rectangles are target objects. Object instances are visualized as small squares colored according to semantic class.
Fig. 3 : Performance results in the primary target search task: Comparison of our method to oracle methods, displaying the episode-wise path length ratio (PLR) to the coverage baseline (dashed line) as boxplots.
Method
SPL (Task Setup 1)
SPL (Task Setup 2)
Coverage Priorities
0.406 ± 0.196
0.341 ± 0.313
Oracle Priorities
0.704 ± 0.202
0.564 ± 0.275
Oracle Intervention
0.712 ± 0.206
0.529 ± 0.281
Linear Oracle Priorities
0.650 ± 0.207
0.520 ± 0.271
Learned Priorities (ours)
0.627 ± 0.225
0.520 ± 0.313
TABLE I : Comparison of our method with the coverage baseline and oracle methods using the SPL metric defined in Section V-B1 , given as mean ± std.
Fig. 4 : Comparison of different dataset sizes and oracle behaviors used for training, displaying PLR performance of the resulting priority models.
Fig. 5 : Top-down views of an example scenario of the second task setup comparing the behavior of coverage-driven exploration and our learned semantic priority model. Visuals follow the same conventions as in Fig. 2 .
Fig. 6 : Performance results in the secondary target search task: Comparison of our method to oracle methods, displaying the episode-wise path length ratio (PLR) to the coverage baseline as boxplots.