The authors evaluated 18 open VLA models spanning four action-generation strategies (flow matching, parallel decoding, token-by-token decoding, and reasoning-before-action) on two architecturally distinct edge platforms: the NVIDIA Jetson AGX Thor and the AMD Ryzen AI Max+ PRO 395.
VLA model latency on edge depends more on action-generation strategy than size.
Benchmarking 18 models with a seven-stage inference breakdown reveals that kernel fusion and few-step generation are key to real-time robot control.
Top University
Borui Xu · Hongshi Tan · Wenqi Jiang · Bingsheng He · Ming Wu · Jason Zeng · +1 more
National University of Singapore · 0G Labs
Research Digest··2 min read
The authors benchmarked 18 open VLA models on two edge platforms (NVIDIA Jetson AGX Thor and AMD Ryzen AI Max+ PRO 395), breaking each inference call into seven stages.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.