OpenBMB Releases MiniCPM-Robot Models and PhyAI Inference Framework
OpenBMB has open-sourced MiniCPM-Robot, a series of lightweight embodied AI models, which includes the 1.5B MiniCPM-RobotManip for robotic manipulation and the 0.5B MiniCPM-RobotTrack for target tracking. Alongside these models, they released PhyAI, a high-performance inference framework designed specifically for physical AI. Providing lightweight, open-source Vision-Language-Action (VLA) models enables more accessible, edge-based robotics and embodied AI applications. This lowers the barrier to entry for researchers and developers looking to deploy physical AI models directly on-device without relying on heavy cloud infrastructure. The MiniCPM-RobotManip model acts as a general-purpose controller for robotic manipulation, while the smaller MiniCPM-RobotTrack is optimized for real-world tracking. The accompanying PhyAI framework supports both cloud-based serving and on-device deployment for these physical AI models.
## BACKGROUND
Embodied AI refers to artificial intelligence systems that interact directly with the physical world through sensors and actuators, rather than operating solely in digital environments. Vision-Language-Action (VLA) models are a class of multimodal foundation models that translate visual inputs and textual instructions directly into low-level robotic actions.