PhyAI: Real-Time Physical AI at the Edge, Scalable Rollouts in the Cloud
PhyAI is a Physical AI inference engine that unifies inference across model evaluation, cloud reinforcement learning rollout, edge GPU serving, and onboard deployment using a single runtime. It supports vision-language-action (VLA) models and world-action models (WAMs) on single or multiple GPUs. PhyAI achieves 1.40x-4.65x speedups over official implementations of pi0, pi0.5, GR00T N1.7, and MiniCPM-Robot. On Cosmos3-Nano-Policy-DROID, it reduces latency from 2.46 to 1.18 seconds on eight H20 GPUs (CFG=2, TP=4), a 2.08x speedup. The adapter interface allowed adding MiniCPM-Robot on its release day.
PhyAI separates architecture-specific conditioning, solver, cache, and output logic into model adapters while sharing graph execution, kernels, memory management, and parallel services. This design enables a single codebase to run diverse physical AI models across deployment scales. The reported speedups indicate efficient utilization of GPU resources, though specialized runtimes may still outperform in some configurations.
Unifying inference runtimes for physical AI can reduce engineering overhead and accelerate deployment cycles for robotics and embodied AI companies. The ability to quickly integrate new models like MiniCPM-Robot suggests a flexible platform that could lower barriers to adopting state-of-the-art physical AI policies.
PhyAI's unified runtime can streamline the development and deployment of physical AI applications, potentially reducing costs and time-to-market for robotics and autonomous systems. Its performance gains may enable more complex real-time inference at the edge, expanding the feasible use cases for physical AI.
Observable next signals include adoption of PhyAI by robotics platforms, integration with additional physical AI models, and benchmarks comparing its performance against specialized runtimes in more diverse hardware setups. Further speed improvements and support for emerging model architectures are likely.