Tsinghua's PhyAI engine unifies physical AI inference across cloud and edge

jiqizhixin · x · 2026-08-22

Researchers from Tsinghua, Peking, and BUPT introduced PhyAI, a unified inference engine for physical AI. It allows the same code to run across cloud GPUs, edge servers, and onboard robots without rewriting. Benchmarks show 1.40×–4.65× speedups over leading VLA models like GR00T N1.7 and MiniCPM-Robot. On an 8-GPU setup, latency for the Cosmos3 DROID policy dropped from 2.46s to 1.18s.

Original post →

More from Infra

Infra channel →