On July 19, OpenBMB (Mingxi Intelligent) open-sourced MiniCPM-Robot, the first series of embodied AI models designed to make vision-language-action (VLA) technology accessible to individual developers. The release includes three core components: MiniCPM-RobotManip, a 1.5B parameter VLA model for general robot manipulation; MiniCPM-RobotTrack, a compact model for real-world target tracking; and PhyAI, a high-performance inference framework built specifically for embodied models.
Until now, embodied intelligence has been largely confined to large corporations and research laboratories with access to massive compute resources and expensive robot hardware. The MiniCPM-Robot series breaks this barrier by enabling personal developers to run real robot operations and target tracking using the 1.5B parameter models, which can be deployed on consumer-grade hardware.
MiniCPM-RobotManip, the flagship model of the series, is a general-purpose VLA model that translates visual observations and language instructions into precise robot actions. Unlike larger models that require data center-class GPUs, the 1.5B parameter scale allows the model to run on a single consumer GPU while maintaining competitive performance on manipulation tasks including pick-and-place, assembly, and tool use.
MiniCPM-RobotTrack addresses a complementary challenge: enabling robots to track and follow targets in dynamic real-world environments. The compact model is optimized for real-time inference, making it suitable for applications such as autonomous navigation, human-robot interaction, and mobile manipulation scenarios.
PhyAI, the accompanying inference framework, is designed from the ground up for embodied AI workloads. It provides optimized implementations of key operations including visual feature extraction, action decoding, and trajectory optimization, with specific attention to the latency requirements of real-time robot control.
All model weights and code are publicly available on GitHub (github.com/OpenBMB/MiniCPM-Robot), with separate downloads for MiniCPM-RobotManip and MiniCPM-RobotTrack on Hugging Face. The open-source release marks a significant step toward democratizing embodied AI research and development.
