Mingxi Intelligent Open Sources MiniCPM-Robot: Full Release of a 1.5B VLA Model, Personal Developers Can Now Run Operations on Real Robots

For the past, embodied intelligence was basically a game for big companies and laboratories—the models were too large and the barriers too high, and ordinary developers couldn't even touch real robots. On July 19, OpenBMB kicked open this door. It released the first series of embodied AI models, MiniCPM-Robot, fully open-sourcing a 1.5B parameter vision-language-action model, allowing individual developers to run operations and target tracking on real robots. This level of openness is rare in the field of embodied intelligence.

This series includes three core components. One of the main ones is MiniCPM-RobotManip, a 1.5B parameter general vision-language-action (VLA) model, specifically responsible for robot manipulation. The other is MiniCPM-RobotTrack, a compact model for real-world target tracking. In addition, OpenBMB also released PhyAI, a high-performance inference framework designed specifically for embodied models. Its goal is clear: to enable robots to truly understand, remember, and act. Together, these three components bring efficient, practical, and open embodied intelligence from papers and demonstrations into real-world robots.

For developers, the most practical thing is that it's easily accessible. Model weights and code are hosted in the GitHub repository github.com/OpenBMB/MiniCPM-Robot, and the weights of MiniCPM-RobotManip and MiniCPM-RobotTrack are also separately available for download on Hugging Face. Those working on robots can now clone and run them directly.

When the 1.5B VLA model is no longer locked in the lab, the pace of embodied intelligence may exceed many people's expectations.

Project address: https://github.com/OpenBMB/MiniCPM-Robot

Wang Xingxing from HuaZhu says at the World Internet Conference: The 'ChatGPT Moment' for Humanoid Robots Is Coming in as Few as Two or Three Years

Humanoid robots have gone from stumbling to dancing and fighting in just a few years, and the critical point from being able to move to being able to work is approaching rapidly. On July 22nd, at the Digital Silk Road Development Forum held in Xi'an, 2026 World Internet Conference, Wang Xingxing, CEO of HuaZhu Technology, proposed an aggressive timetable: the 'ChatGPT Moment' of embodied intelligence will arrive as early as two or three years from now, at which time robots will truly have the capability to work and overcome the high barrier to practical application.

AntGroup's LingBot-Video Open Source: The World's First Video Foundation Model for Embodied Intelligence!

AntGroup opensources the world's first video foundation model for embodied intelligence, LingBot-Video. This model restructures pre-training around robot requirements, systematically improving reasoning efficiency, physical plausibility, action understanding, and task completion. It provides an open-source foundation for embodied intelligence and has been validated on the RBench benchmark jointly released by Peking University and ByteDance.

New Breakthrough in Embodied Intelligence: Ant Group Open Sources LingBot-Vision, Enabling Robots to Have a Sense of Space

Ant Group's Robbyant opensources the LingBot-Vision model family, which achieves outstanding performance in dense space perception tasks through self-supervised vision Transformers and innovative boundary modeling. It surpasses large models with several times more parameters in multiple metrics, breaking the limitations of existing visual foundation models that focus heavily on object recognition, making precise perception of physical space by robots a reality.

The 'ChatGPT Moment' for Embodied Intelligence is Approaching: Yuque Technology Envisions a New Landscape for the Robot Industry

Chen Li of Unitree Robotics stated at the 2026 Yabuli Forum that embodied intelligence is on the verge of an explosion, with breakthroughs expected in 2–5 years. Its 'ChatGPT moment' will be marked by robots completing ~80% of tasks accurately in 80% of unfamiliar scenarios via voice/text commands. He predicts this milestone is near.....

Robots Gain a General Brain: RoboScience Releases Visics Large Model, Achieving Cross-Scenario Autonomous Execution

Robots are moving towards general embodied intelligence. RoboScience released the large model Visics, which uses the VLOA architecture, breaking the repetitive training mode of single tasks and achieving universal operation capabilities across different bodies, objects, and tasks, leading the industry to move away from the 'action replication' fragmented era.