Two engineers built a DIY robot vacuum cleaner for under $300 using off-the-shelf hardware and a behavior cloning approach for autonomous navigation. They teleoperated the robot to collect image-action training pairs, then trained a CNN to replicate those navigation decisions. The model struggled with key tasks like detecting free space and handling rotational commands, largely due to insufficient signal in the dataset. Experiments with data augmentation and ImageNet pre-training confirmed the issue was data quality rather than overfitting. The project took about 4 months part-time and identified clear areas for improvement: better data collection, adding temporal context via frame history, and stronger vacuum hardware.
Sort: