VISUAL AI PROTOTYPE
2,351,952
parameters in the current architecture
Camera + sensorsTime + context
Architecture prototype; not a released navigation model yet.
IN DEVELOPMENT · SHADOW ONLY
A second
pair of eyes
on the water.
AI BoatPilot is an independent development project for perception on the water. We explore how AI, camera footage and onboard data can help us understand the water around a boat.
A Raspberry Pi gathers observations from the front camera, GPS and NMEA 2000. We investigate and evaluate models on our own server, learning from real-world conditions one step at a time.
At this stage the software only observes: it does not control the boat. The skipper remains at the helm. This is an experimental project, not a certified navigation system.
View live model status ↘ Participant sign-in ↗THE PRINCIPLE SCHEMATIC
↑
OBSERVE · UNDERSTAND · COMPAREFRONT CAMERA + N2K / GPSNO STEERING COMMANDS
THE TECHNOLOGY BEHIND THE PROJECT
Follow the development.
AUTOMATIC IMAGE TRAINING
Status temporarily unavailable
New valid front-camera images automatically start self-supervised pretraining. Object detection trains when reviewed labels are available.
— recordings prepared · — image-learning runs · — object-learning runs
MODEL WEIGHTS
8.97 MiB
Calculated for FP32 weights. Total memory needed during training is higher.
TRAINING PIPELINE
Standing by for the next round
Synthetic test data · CPU sensor baseline
6 weights in the training baseline. This is a separate pipeline test.
COMPLETED EXPERIMENTS
54
Completed baseline runs on the server. Not a navigation safety or recognition accuracy score.
MODEL FOR THE PI
Not released yet
Release requires evaluation and a real hardware test. The software remains in shadow mode.
AI tools have been used to develop the software and draft the text. The detailed project journal and participant area currently require sign-in.