Robota: Companion robot
A concept implementation of a companion robot utilizing scene recognition. Took an excelence award in a Hackathon hosted by Yahoo.
Preview
Details
A concept implementation of an everyday robot companion "Robota".
Robota will wake you up, greet you when you come back home, praise you when you work hard, scold you when you sleep too long etc.
A webcam is placed in the room to capture the sitution and allow for these kinds of interactions.
Worked as a team of two. I was in charge of the action estimation, my friend of the robot controls.
At first, precision of the action estimation was very low. This was due to overfitting since we only had less than a thousand images. The solution was to use a larger variety of webcam angles, in combination with more aggressive data augmentation.
System overview
Real-time action estimation in a custom environment
Presentation Video from Hackathon (Japanese, 180s)
Tools
Trained with pytorch.
Backbone Model is MobilenetV2, since this had to run in real-time on a laptop.
Weights & Biases was used to monitor all training runs.
Docker was used to ensure the same environent between training (A multi-GPU server) and running for inference (a laptop)