Effective emergency response hinges on rapid, precise situational awareness. While AI modules currently exist to analyze UAV and satellite imagery, there remains a critical void in ground-level perception for post-disaster environments. To bridge this gap, we present a pioneering perception software solution designed specifically for Unmanned Ground Vehicles. Our tool serves as an advanced intelligence engine, enabling the rapid deployment of diverse robotic platforms into active Disaster Response Systems. By shifting the focus to the ground level, we provide first responders and autonomous systems with the granular, up-close intelligence necessary to execute precise hazard localization, boundary delineation, and severity estimation. The core of our product is the first tool capable of ground-level disaster AI recognition, comprehensively categorizing hazards into 53 unique, task-specific classes grouped under 5 major categories. Unlike conventional resources that offer simple image categorization without spatial context, our solution delivers dense, pixel-level environmental understanding. It empowers models to differentiate specific hazards from background noise, granting ground robots the semantic intelligence required to operate autonomously. Powered by a rapidly expanding foundation of operational data, our perception module is designed to integrate seamlessly into next-generation DRS architectures. Currently at TRL 5, our technology is actively validated using real-world operational imagery. The system is explicitly engineered to thrive in cluttered, highly dynamic, and exceptionally harsh environments. Rigorous benchmarking demonstrates reliable hazard boundary extraction and autonomous navigation support even under extreme conditions, including severe occlusion, dense smoke, fire, debris, and dust. By integrating our baseline models into a prototype interface, we are rapidly advancing to TRL 6.
Contact
The primary objective of TerraRescue is to revolutionize ground-level disaster response by equipping UGVs with advanced autonomous perception. Currently, first responders face immense risks in volatile, unmapped environments. Our software transforms standard robots into intelligent frontline assets that autonomously navigate through chaotic hazardous zones to deliver real-time, actionable intelligence. The technology facilitates autonomous safe-path detection, precise hazard localization, and structural severity estimation. By grasping exact spatial boundaries instead of merely classifying scenes, the UGV can actively maneuver around obstacles, assess damage, and operate safely in dynamic, smoke-filled environments without continuous human intervention. To achieve this level of operational readiness, the robot’s perception engine is powered by a highly specialized semantic corpus. The system is trained to instantly recognize and spatially locate 53 unique, incident-specific classes grouped in 5 major operational categories. These are: a) General-purpose (e.g., Citizen, Green/Dry/Burned tree, etc.), b) Infrastructures (e.g., Building, Pole, Window, etc.), c) Disasters (e.g., Smoke, Destroyed building/vehicle, etc.), d) First Responder vehicles (e.g., Ambulance, Police Vehicle, Fire Truck, etc.), and e) PPE & other equipment (e.g., Helmet, Extinguisher, Fire hydrant, etc.). This extensive visual vocabulary allows the UGV to effectively parse complex disaster scenes, identifying active hazards. Furthermore, it can identify deployed first responder vehicles and crucial protective equipment, including SCBA units, helmets, and fire hoses. Ultimately, our ambition is to transition ground-level disaster AI from a conceptual phase into a deployable, life-saving tool. By granting ground robots high-fidelity spatial understanding under authentic, harsh conditions, we provide an autonomous solution that directly accelerates emergency response times.
The true innovation of our solution lies in its unprecedented ability to bring highly granular, autonomous AI perception directly to the ground level of active disaster zones. While most existing technologies rely heavily on high-altitude, low-resolution aerial overviews or ground-level analysis with very limited semantic corpus capabilities, our solution equips UGVs with immediate, human-like spatial awareness in the most chaotic and unpredictable environments. This transforms UGVs from remotely piloted cameras into intelligent, context-sensitive partners capable of proactive decision-support. Our system moves far beyond basic obstacle avoidance by utilizing an advanced perception engine that instantly parses complex scenes into 53 disaster-specific classes across five major categories: general purpose, infrastructures, disasters, first responder vehicles, and personal protective equipment. This comprehensive visual vocabulary allows a UGV to dynamically distinguish between navigable areas and impassable hazards like mud. It can also identify structural debris or a hole in the ground, adapting its autonomous pathing in real-time to ensure safe and efficient navigation. Furthermore, the software transitions disaster AI from simple categorization to dense, pixel-perfect spatial understanding. It explicitly delineates the boundaries of active threats, such as flame and smoke, against background noise. The system also recognizes critical operational elements, differentiating a first responder from a citizen, and identifies vital equipment, such as a fire hose or an SCBA unit. By anchoring our algorithms in operational realism rather than synthetic simulations, our perception engine remains exceptionally robust. It maintains high-fidelity hazard localization even under severe occlusion, unpredictable lighting, and the highly dynamic physical conditions typical of post-disaster environments.
TerraRescue addresses a critical void in current emergency response: the lack of granular, ground-level situational awareness for autonomous systems in post-disaster environments. While aerial imagery provides broad overviews, first responders currently face immense risks navigating volatile, unmapped zones where high-altitude data cannot capture the "on-the-ground" complexities. Our solution resolves this by transforming standard Unmanned Ground Vehicles (UGVs) into intelligent, autonomous assets capable of human-like spatial awareness in chaotic scenarios. The core engine overcomes three primary operational hurdles: 1) Semantic Intelligence Overload: Traditional AI often fails in the visual complexity of disaster zones. TerraRescue utilizes a specialized semantic corpus of 53 task-specific classes (across 5 categories: General, Infrastructure, Disasters, First Responders, and PPE) to identify everything from "Burned trees" to "SCBA units" and "Fire hoses". 2) Navigational Uncertainty: Unlike systems that only provide simple image categorization, our tool delivers pixel-level environmental understanding. This allows UGVs to differentiate between navigable surfaces and impassable hazards, such as distinguishing "Water" from "Mud" or identifying "Structural debris" and "Holes" in real-time. 3) Environmental Extremes: The software is built to thrive where human perception fails. It maintains reliable hazard boundary extraction and autonomous pathing under extreme conditions, including dense smoke, fire, severe occlusion, and dust. Ultimately, TerraRescue resolves the "blindness" of ground platforms, enabling precise hazard localization and structural severity estimation without continuous human intervention. This directly accelerates response times and significantly reduces the risk to human life by allowing robots to lead the way into the most unforgiving environments.
Led by Asst. Prof. Georgios Th. Papadopoulos (Harokopio University of Athens), our team features 2 PhD researchers and 15 diverse data and backend engineers specializing in deep learning and visual scene perception. Combining deep expertise in computational vision with a scalable, custom backend infrastructure, we deploy advanced models to drive the rapid, high-quality development of our pioneering ground-level perception software.
Led by Asst. Prof. Georgios Th. Papadopoulos (Harokopio University of Athens), our team features 2 PhD researchers and 15 diverse data and backend engineers specializing in deep learning and visual scene perception. Combining deep expertise in computational vision with a scalable, custom backend infrastructure, we deploy advanced models to drive the rapid, high-quality development of our pioneering ground-level perception software.