Open-Source Embodied AI for Real-World Robots
OpenBMB has officially launched the MiniCPM-Robot series, a new family of open-source embodied intelligence models engineered to bring high-speed, on-device AI capabilities to real-world robotic systems. The release, made available under the Apache-2.0 license on both GitHub and Hugging Face, represents a significant step toward making advanced vision-language-action (VLA) technology accessible to robotics developers without requiring massive compute infrastructure.
The announcement addresses a long-standing challenge in embodied AI: while large-scale vision-language models have demonstrated impressive reasoning capabilities, their size and inference latency make them impractical for direct deployment on robotic hardware that demands real-time responsiveness.
Two Specialized Models, One Unified Framework
The MiniCPM-Robot series comprises two specialized models covering complementary aspects of embodied intelligence, plus a dedicated inference engine optimized for throughput:
- MiniCPM-RobotManip (1.5B) — A general-purpose Vision-Language-Action model for robotic manipulation tasks. Using a unified policy with a single set of weights, it handles diverse downstream tasks across both simulation and real-world environments. Despite its compact size, it outperforms larger models including π₀.₅ (3B) and Qwen-VLA (5B+) on representative evaluation benchmarks.
- MiniCPM-RobotTrack (0.9B) — An ultra-compact model optimized for embodied target tracking. It integrates natural-language instructions with fused DINOv3 and SigLIP visual features to predict future waypoints for tasks such as person following, achieving SOTA results among open-source models on the EVT-Bench benchmark.
- PhyAI Inference Engine — A high-performance runtime framework providing Day-0 support for the series. On NVIDIA H20 hardware, PhyAI increases inference throughput from 10 Hz to 37 Hz using CUDA Graph and custom Triton fused kernels — a 3.7× speedup that is critical for real-time robotic control loops.
Technical Innovations Driving Performance
MiniCPM-RobotManip implements streaming inference that incorporates historical observations into the model context, allowing it to maintain up to one minute of visual memory while significantly reducing computational overhead. The model also inherits visual token compression from MiniCPM-V 4.6, reducing the frame footprint from 256 to 64 visual tokens — a 4× compression ratio that directly translates to faster inference and lower memory usage.
For tracking, MiniCPM-RobotTrack leverages a quality-driven self-evolving data pipeline combined with DAgger-style model-environment interactions to improve generalization in complex real-world scenarios including target crossings, rapid turns, and partial occlusions.
On H100 hardware using BF16 precision, RobotManip achieves a forward latency of just 120 ms per decision step — well within the 200 ms threshold generally considered necessary for responsive robotic manipulation. Deployed on a Unitree Go2 EDU platform, RobotTrack maintains a stable 5+ FPS with approximately 180 ms end-to-end latency using fully local, vision-only natural-language tracking — no cloud dependency required.
Why This Matters for the Embodied AI Ecosystem
The release signals a maturing ecosystem where specialized, efficient models are beginning to outperform one-size-fits-all large models on specific robotic tasks. By open-sourcing both the models and the inference framework, OpenBMB is lowering the barrier to entry for robotics developers who need real-time on-device intelligence but lack the resources to train or deploy massive foundation models.
The trend toward smaller, specialized models has important implications for humanoid robotics. As robots move from controlled laboratory environments into unpredictable factory floors, warehouses, and homes, the ability to make competent decisions locally — with minimal latency and no reliance on cloud connectivity — becomes a prerequisite for reliable and safe deployment.
OpenBMB, a Chinese AI research organization best known for its MiniCPM series of compact language models, is positioning itself as a key player in the embodied AI space by extending its expertise in efficient model design from the digital domain into the physical world.
