Developer resource system building
Xiaomi has built a comprehensive open source support system for Orion:
- Pre-trained model library: contains core components such as 2D language model weights (Hugging Face), visual coder (OmniDrive), etc.
- Data preprocessing tool: supports automatic parsing and format conversion of Bench2Drive datasets
- Visual analytics suite: comparative videos with decision markers can be generated (marking key nodes such as traffic lights, obstacle interactions, etc.)
- Evaluation indicator system: integration of 12 specialized indicators such as L2 error, Jerk value, lane keeping rate, etc.
The technical architecture is modular and the main interfaces include:
- Scene inference API: process_scene(image, history)
- Trajectory generation API: generate_trajectory(instruction)
- Evaluation interface: eval_loop(dataset, metrics)
The development documentation details secondary development methods for each module, such as implementing custom planning algorithms by inheriting the BasePlanner class. The community version of the roadmap indicates that the training framework code will be released in Q3 2024.
This answer comes from the articleOrion: Xiaomi's Open Source End-to-End Autonomous Driving Reasoning and Planning FrameworkThe
May not be reproduced without permission:AI productivity tools " Orion's open source ecosystem includes a complete research and development tool chain
