Zhiyuan has made a significant technological leap in the on-device deployment of embodied intelligent VLA (Vision-Language-Action) models. By employing a combined approach of algorithmic and engineering optimization, Zhiyuan has dramatically enhanced the inference frame rate of the π0.5 VLA model on the NVIDIA Jetson Thor platform. The frame rate has soared from a mere 1.4Hz to an impressive 22.1Hz, marking a staggering performance boost of over 15 times.

This optimization effort cuts across multiple technical strata, empowering robots to break free from the shackles of external desktop graphics card computation. It enables seamless, on-device inference on the Jetson Thor chip, a feat that has been rigorously tested and verified on real-world machines, exemplified by the Elves G2 robot.

This groundbreaking achievement is geared towards cost reduction and performance enhancement. The optimized frame rate not only outstrips the current best industry benchmarks but also stands as a pivotal exploration in propelling the widespread adoption of embodied intelligence.