SteerEfficient autonomous vehicle software |
Version | 0.1 | |
|---|---|---|---|
| Source | repo | ||
The Transformer architecture has become widely used for all sorts of tasks, such as language, vision, genetics and many more. I think attention ain't all we need. I also think there isn't a universal architecture we have to find, the answer can be in the balance. That's why I am looking forward to develop hybrid models, based on attention and state spaces. The first Steer model is using the Video Mamba architecture, with a focus on performance at inference time.
Why self-driving? it's a hard task, requiring multiple data inputs, local and global temporal awareness, complex data relationships. This project does not aim to provide full autonomy, but to test the capabilities of our architecture. We also don't have the necessary hardware to train such big models, yet.
Read more about the research process, model architecture and findings in the tehnical report: english and romanian.
I'm Stefan Asandei, an 11th grade student from Romania. I build open software on github and write about it on my blog.
November 2nd, 2024