🚀 MiniCPM enters the physical world — enabling robots to understand, remember, and act. We open-source MiniCPM-Robot, our first embodied AI model series, including: 🤖 MiniCPM-RobotManip — a 1.5B general-purpose Vision-Language-Action (VLA) model for robotic manipulation. 🐕 MiniCPM-RobotTrack — a compact model for real-world target tracking. ⚡ PhyAI —... a high-performance inference framework built for embodied models. Together, they bring efficient, practical, and open embodied intelligence closer to real-world robots. ⭐ GitHub: 🤗 MiniCPM-RobotManip: 🤗 MiniCPM-RobotTrack:show more

Wishing the team continued momentum.

A robot doesn't just need to see what is happening. It needs to remember what happened before. Imagine asking a robot: "Press the second button five times." Without contextual memory, it may lose track of previous actions and fail to complete the task. MiniCPM-RobotManip (1.5B) tackles this challenge with highly efficient visual token compression inherited from MiniCPM-V 4.6, enabling streaming native memory context without additional inference cost. As a general-purpose VLA model, it achieves performance comparable to leading models such as π₀.₅ (3B) and Qwen-VLA (5B+), while achieving SOTA performance on memory-intensive robotic benchmarks with 1.5B parameters. Smaller. Faster. Better memory.

Robots shouldn't stop working when the network does. Real-world environments are unpredictable: people move around, targets disappear, and networks become unstable. MiniCPM-RobotTrack is a compact Vision-Language-Action model built on MiniCPM4-0.5B for robust target tracking in real environments. It supports: ✅ Zero-shot language-guided tracking ✅ Robust tracking in dynamic multi-target environments ✅ Ambiguous instructions and challenging scenarios ✅ Fully local deployment without cloud dependency Native support for Unitree Go2 Edu enables robots to understand instructions and follow targets using only onboard vision and local computation. Even in weak-network or offline environments such as elevators and underground parking lots, robots can continue operating reliably.

Great embodied models also need efficient inference. That's why we're introducing PhyAI, an open-source inference framework designed for Physical AI. Built for both cloud-based serving and on-device deployment, PhyAI makes it easier to bring embodied models from research into real-world applications. With CUDA Graph optimization and custom Triton fused kernels, MiniCPM-Robot inference throughput increases from 10 Hz to 33 Hz, and further to 36 Hz on NVIDIA H20. As Physical AI models evolve rapidly, we hope PhyAI helps the community build faster, more efficient embodied intelligence together.

Embodied AI will be built by an open community. Explore, build, and create with MiniCPM-Robot: ⭐ GitHub: 🤗 MiniCPM-RobotManip: 🤗 MiniCPM-RobotTrack: If you build something with MiniCPM-Robot, we'd love to see it. Share your projects, feedback, and ideas with the community.

Loving the offline tracking capabilities on Unitree Go2 look super practical for everyday robotics.

Impressive milestone...... Great to see efficient embodied AI becoming open source. Looking forward to what the community builds.

Curious how much of this transfers to new objects without task-specific tuning

It's refreshing to see a release focused on practical deployment instead of just benchmark numbers.

The next challenge is handling long multi-step tasks without forgetting earlier actions

Closing that generalization gap is exactly what general-purpose VLA models should be solving.

Local inference seems to be the right direction for robots that operate near people.

I'd love to try this on my own setup.

This is awesome. Looking forward to seeing more MiniCPM-Robot demos at WAIC

Big week for open-source embodied AI. Looking forward to what's next.

Robots that understand, remember, and act is a big step forward.

Great to see more open source robotics work

Huge step for embodied AI.

Open-source embodied AI is exactly what the ecosystem needs right now. Looking forward to trying this out.

@AdinaYakup Thank you ❤️🎉🎊 another excellent release

@Presidentlin

solid step toward on‑device embodied ai. contextual memory will make multi‑step tasks feasible

@grok is there a benchmark test for humanoid robots like the Frontier AI model testing? Tell me more. Thanks

vla quality matters, but recovery policy after bad grasps decides production value

Great contribution to the open AI ecosystem. MiniCPM-Robot, together with PhyAI, can become a reference for those who work in robots that are more efficient, accessible and capable of operating in the real world.

Open source embodied AI models, big for robotics research

Does this work out of the box?

It’s designed for easier deployment, but it still requires integration with a supported robotic setup.😍

Will there be support for custom robot hardware?

Would love to see an uncut demo with a few failed attempts included....that would make the real-world reliability much easier to judge.

Curious how this holds up outside the demo.

Team China moves mountains.

Love the quality! 💯 Great job.

1.5B model making robots actually understand and act open source

It's exciting to see foundation-model teams bringing their efficiency work into robotics

The offline part is insane bro

Great step toward making embodied AI more open and practical

Thanks! We’re excited to contribute to a more open and practical future for embodied AI.

Would be interesting to see a raw demo.