All work
ARUW — Real-Time Perception
Optimized a Jetson-based detection and classification pipeline for competition robotics.
Detection pipeline on the Jetson. Select a stage.
The problem
Competition robots need perception results quickly enough to act on moving opponents. The whole pipeline matters: preprocessing, model execution, postprocessing, and moving data between CPU and GPU.
My contribution
I optimized a YOLO-based detector and classifier on NVIDIA Jetson using TensorRT, CUDA, quantization, and GPU-accelerated preprocessing and postprocessing. I also improved model execution and CPU–GPU memory transfers.
Result
End-to-end latency fell from roughly 30 ms to 6–7 ms. ARUW placed third at the 2026 ARC Robotics Competition, a team result.