Robot model study reports 2.6-2.9× inference speedups on Jetson hardware

Researchers report in the VLA-HWBench preprint that combining fused computing operations with 4-bit model weights ran π₀.₅ and GR00T N1.7 2.6-2.9× faster than PyTorch on NVIDIA Jetson hardware. The study compares 18 open models that turn visual and language inputs into robot actions across NVIDIA Jetson AGX Thor and AMD Ryzen AI Max+ PRO 395 hardware. Action-generation strategy mattered more than model size: parameter count explained at most 15% of latency variance. Models producing long action sequences in a few steps kept up with control, while token-by-token decoding did not. The accelerated implementation exists only for those two models on Jetson; CPU preprocessing rose from 12% to 39% of GR00T N1.7's processing time.