Ing. Tomáš Báča, Ph.D.

The operational ceiling for autonomous aerial systems is defined by their ability to perceive and navigate complex, cluttered environments where external positioning systems like GPS are unavailable. Current approaches often falter due to their reliance on unimodal sensing, which is inherently fragile—vision fails in poor light, while LiDAR struggles with textureless surfaces or adverse weather. While multimodal systems offer a path forward, they typically fuse data at a purely geometric level, creating a significant computational bottleneck that is often untenable for the limited size, weight, and power constraints of an onboard aerial platform. This research addresses this critical gap by proposing a paradigm shift from geometric-level fusion to onboard semantic fusion. Instead of merely merging raw sensor data, this project will investigate how to extract high-level, semantic meaning from disparate data streams (e.g., RGB cameras, LiDAR, thermal) and fuse this understanding into a unified, lightweight, and robust world model. By reasoning about the environment in terms of recognized objects and structures (e.g., "trees", "walls", "pathways"), the system can achieve a holistic perception that is more resilient to the failure or ambiguity of any single sensor. We identify three core issues that will have to be investigated: (i) Efficient Cross-Modal Semantic Association: The foundational challenge is to develop lightweight algorithms capable of extracting semantic features from different sensor modalities and, crucially, associating them in real-time. This involves answering the question: How can the system determine that a specific group of pixels in a camera image corresponds to the same physical "doorway" represented by a void in a LiDAR point cloud?; (ii) A Novel Semantic Fusion Engine: The central contribution will be the design of a probabilistic fusion framework that runs entirely onboard the aerial platform. This engine will take the associated semantic features as input and generate a single, coherent estimate of the environment's state. It must be designed for robustness, capable of gracefully handling sensor dropouts and uncertainty to produce a continuous and reliable perception output; (iii) Semantically-Informed Localization and Navigation: The ultimate goal is to leverage the fused semantic model for superior performance in navigation tasks. This research will explore how the system can use recognized semantic landmarks for highly robust localization (e.g., "I am near the fire hydrant") and for context-aware path planning (e.g., "plan a path through the corridor while avoiding the movable chairs"), leading to safer and more intelligent autonomous behavior in cluttered spaces. By pioneering an onboard semantic fusion framework, this research aims to unlock a new level of robustness and intelligence for aerial robots, enabling them to navigate reliably and autonomously in the challenging, real-world environments where they are needed most.