Project #21 - Optimize a VLA model for embodied intelligence

[jayden1711]

Hello Na Cui, Baihe Liu!

I'm Jayden, a sophomore CS student at Ohio State interested in this project. I spent the last year optimizing a DinoV2 ViT for real-time deployment on a Jetson AGX Orin, cutting latency 50% through CUDA, cuDNN, and TensorRT tuning.

I think the hardware optimization work here translates directly, even switching from CUDA to oneDNN. Just to give an overview of what my initial thoughts of this project are, I would most likely approach this by profiling the target VLA model first to find bottlenecks, then systematically applying quantization and kernel-level optimizations while benchmarking at each step.

The Good First Issues board is fully claimed, so could you assign me a prerequisite task directly? I also thought any tasks you point me to would probably be more applicable to the project.

Thanks!