What is minicpm-robot?

MiniCPM-Robot is a family of small AI models built to give robots the ability to see, decide, and act in the physical world, running directly on the robot's own hardware rather than depending on a distant server. It comes from the team behind the MiniCPM line of compact AI models, and this release focuses on giving robots a brain that fits on device. The family currently includes two models. The first, called MiniCPM-RobotManip, is a general purpose model for robot arms and manipulation tasks, meaning picking things up, moving objects, and similar physical actions. A single version of this model can handle many different tasks rather than needing a separate model retrained for each one, and the makers report it performs better on standard tests than some larger, more resource hungry AI models built by others. It is also designed to remember roughly a minute of recent visual history so a robot's actions can take recent context into account, while still responding about as quickly as a version of the model that only looks at the current moment. The second model, MiniCPM-RobotTrack, is built specifically to track a target that a robot is told to follow or watch, described in plain language, using only the robot's camera and no separate cloud service. It handles targets that are still, moving, or even trying to evade tracking, and the makers state it runs fast enough, several frames per second, to work on a real four legged robot in real time. Both models are shared as open source AI models that developers can download and build with themselves rather than access only through a paid service. Getting them running requires technical setup: installing the right software dependencies, downloading the model weights, and connecting them to compatible robot hardware or a simulation environment. This project is aimed at people already building or researching robotics, not a ready made consumer robot product.