EcoFloater
On-device perception for an autonomous water drone.
Problem and constraints
EcoFloater is a water-drone project at Scrapify Ecotech. My work addressed the onboard AI stack for recognizing aquatic debris and supporting navigation. The challenge was to connect visual perception and sensor information with the compute, power, and connectivity constraints of an edge platform.
The vehicle is a team system. My contribution concerns AI, perception, and deployment work, rather than sole ownership of its mechanical design or complete autonomy stack.
My contribution
I worked on computer vision models for aquatic-waste perception, sensor and depth integration, and deployment on NVIDIA Jetson Nano. The work included attention to inference latency and model size on constrained hardware.
Intel RealSense depth sensing was part of the sensor work. Recognition identifies relevant objects, while spatial information supports reasoning about the surrounding scene.
Model and systems work
The work involved YOLO/YOLOv8 detection, vision-transformer approaches, and DeepLabV3+ segmentation. These represent perception methods used or explored, rather than a claim that all models ran together in a single real-time pipeline on the Nano.
Additional work described in my engineering experience includes LLM decision components, mission sequencing, compression, and quantization. The engineering concern was fitting useful model behavior into an on-device system under resource constraints.
Evidence and limitations
The internship certificate supports the role context; it is not a performance evaluation of the autonomous system. No FPS, power-consumption benchmark, deployment count, or autonomous-mission success rate is claimed here. Those properties require measurements of the particular hardware and integrated pipeline.