PhyAI: A Unified Inference Runtime for Embodied AI with up to 4.6x Speedup

机器之心 · wechat · 2026-08-14

To solve the fragmentation of maintaining multiple codebases across different deployment scenarios, a joint research team proposed PhyAI, a unified inference runtime for Physical AI.

The framework seamlessly transitions across offline benchmarks, cloud RL rollouts, edge deployment, and factory MaaS. The research also introduces Control-Time Roofline, an analytical tool revealing that on devices with sufficient compute, simply reducing inference latency doesn't proportionally increase robot control frequency.

Experiments show that PhyAI achieves 1.40x to 4.65x speedup over official implementations in single-request deployments and significantly reduces inference time in cloud RL rollouts.

Original post →

More from Infra

Infra channel →