Skip to main content

MiniCPM-Robot Embodied AI Series

MiniCPM-Robot is MiniCPM's embodied AI model series for real-world perception, decision-making and action. It focuses on robot manipulation and target tracking, with an emphasis on efficient on-device inference.

Series overview​

MiniCPM-Robot currently covers two directions: MiniCPM-RobotManip and MiniCPM-RobotTrack.

MiniCPM-RobotManip A 1.5B general-purpose VLA model for robot manipulation. A single set of weights covers a wide range of downstream manipulation tasks, with support for streaming visual context.

MiniCPM-RobotTrack A 0.9B on-device embodied target-tracking model that covers static, dynamic and adversarial targets, and can be deployed to the Unitree Go2 EDU.

ModelParametersTaskHighlights
MiniCPM-RobotManip1.5BRobot manipulationGeneral-purpose VLA, streaming visual context, one model for many tasks
MiniCPM-RobotTrack0.9BEmbodied target trackingRuns on-device, natural-language tracking, real-robot deployment on Go2
Use caseRecommended modelWhy
Robotic-arm grasping, general manipulation, simulated and real-robot manipulationMiniCPM-RobotManipUses unified weights to cover many manipulation tasks, with multi-view input and action-sequence output.
Quadruped robot following, dynamic target trackingMiniCPM-RobotTrackOptimized for static, dynamic and adversarial target tracking, and comes with an on-device deployment workflow for the Go2 EDU.
Research on long-horizon visual memory and efficient VLA inferenceMiniCPM-RobotManipStreaming inference can reuse historical visual context, reducing the cost of repeated computation.

Quick advice: For robot manipulation tasks, choose MiniCPM-RobotManip; for target tracking and on-device deployment on the Unitree Go2, choose MiniCPM-RobotTrack.

TopicLink
Model documentationMiniCPM-RobotManip / MiniCPM-RobotTrack
Code repositoryOpenBMB/MiniCPM-Robot
MiniCPM-RobotManip model weightsMiniCPM-RobotManip
MiniCPM-RobotTrack model weightsMiniCPM-RobotTrack