MiniCPM-Robot
A family of on-device embodied-AI models for robots: a 1.5B generalist manipulation policy and a 0.9B on-device target tracker, both designed to run locally without a cloud connection.
MiniCPM-RobotManip is a 1.5B-parameter generalist vision-language-action model for robot manipulation, using one set of weights across tasks in both simulation and the real world; the project reports it outperforming larger models such as a 3B Pi-0.5 and a 5B+ Qwen-VLA on representative evaluations. Streaming inference lets it incorporate up to a minute of historical visual context far more cheaply than recomputing from scratch, and it inherits MiniCPM-V's visual token compression to cut per-frame token cost.
MiniCPM-RobotTrack is a 0.9B fully on-device embodied target tracker, described as reaching open-source state-of-the-art on the EVT-Bench benchmark for static, dynamic and adversarial targets, running at over 5 frames per second, about 180 milliseconds per frame, on a Unitree Go2 EDU robot using only local, vision-based natural-language tracking.