Openpilot
The goal of this project is to implement open-source data collection and training pipeline for comma.ai Openpilot. This involves:
- implementing custom comma.ai API for data collection,
- creating high-precision GNSS trajectories with Laika,
- implementing neural network training,
- integrating the trained network into Openpilot.
Organization
The students will pick roles in the team:
- data engineer - data collection and preprocessing trajectories
- web engineer - comma.ai API, Explorer, Cabana
- NN engineer - neural network training, Qualcomm SDK for neural network inference
- device engineer - everything that happens in the device
If there are not enough people, data and web engineer roles can be skipped. In this case the training will be performed on existing comma2k19 dataset.
Schedule
Links
Other useful sources of information:
- Openpilot wiki
- comma.ai blog
- comma.ai Discord
- comma.ai YouTube channel, including COMMA_CON presentations
- George Hotz YouTube channel
Deliverables
- Each student should present one topic to others. The presentation should include live demo.
- All students together should train one neural network model and demonstrate it driving in comma two device. The driving does not have to be perfect. Either self-collected data or comma2k19 dataset can be used for training.
Contact
Autonomous Driving Lab (room 3095)
Tambet Matiisen
tambet.matiisen@ut.ee