OpenBMB Open-Sources MiniCPM-Robot for Offline Embodied AI
OpenBMB officially open-sourced its first embodied AI model series, MiniCPM-Robot, at the WAIC conference, demonstrating a "disconnect but stay connected" capability where devices maintain stable tracking even after unplugging the robot dog's ethernet. Combined with the PhyAI inference framework, the series achieves significant performance improvements, providing a cost-effective open-source solution for edge-side embodied AI deployment.
Confirmed
The MiniCPM-Robot series primarily includes two core models: MiniCPM-RobotManip and MiniCPM-RobotTrack.
- MiniCPM-RobotManip: A 1.5B VLA general-purpose robot manipulation model based on MiniCPM-V 4.6, focusing on high efficiency with a per-step latency of 120ms on H100. According to @iamfakhrealam, through efficient visual token compression, it requires only about half the visual processing compute of its competitors to rival mainstream VLA models like Qwen-VLA and π0.5.
- MiniCPM-RobotTrack: A 0.9B end-to-end vision-language tracking model (based on MiniCPM4-0.5B) natively compatible with Unitree Go2, supporting offline tracking. It runs entirely on local hardware without cloud or internet, featuring an end-to-end latency of about 180ms and a speed exceeding 5Hz.
For performance optimization, deployment relies on PhyAI, an inference framework for Physical AI. Using CUDA Graph optimizations and custom Triton fusion operators, the model's throughput increased from 10Hz to 33Hz, reaching up to 36Hz on the NVIDIA H20.
Why it matters
Most current robotic demos rely heavily on stable Wi-Fi and cloud compute, causing them to crash if the network weakens or tasks run long. MiniCPM-Robot demonstrates exceptional edge deployment efficiency and offline operational capabilities, effectively solving the pain point of limited compute on edge devices through hardware-specific deep optimizations. Furthermore, as @HeyToha observed, the open-source embodied AI stack's focus on contextual memory and local inference holds the potential to integrate with existing reinforcement learning (RL) pipelines, further advancing the large-scale deployment of embodied intelligence.
2026-07-24 ~ 2026-07-25 · 10 related posts
Primary sources
- [source] OpenBMB opens up MiniCPM-Robot as its first embodied AI model family at WAIC — CyberRobooo · 2026-07-24
- [source] OpenBMB launches a 1.5B MiniCPM-RobotManip model with 120 ms step latency on H100 — CyberRobooo · 2026-07-24
- OpenBMB’s 0.9B MiniCPM-RobotTrack brings offline vision-language tracking to Unitree Go2 — CyberRobooo · 2026-07-24
- [source] PhyAI lifts H20 inference throughput from 10 Hz to 36 Hz for Physical AI workloads — CyberRobooo · 2026-07-24
- OpenBMB Open-Sources MiniCPM-Robot: Embodied AI That Keeps Tracking Offline — iamfakhrealam · 2026-07-24
- MiniCPM-RobotTrack runs on-device at 5+ Hz and keeps tracking unplugged — iamfakhrealam · 2026-07-24
- MiniCPM-RobotManip is a 1.5B VLA that claims half the vision compute — iamfakhrealam · 2026-07-24
- OpenBMB says MiniCPM-Robot jumps from 10 Hz to 36 Hz with CUDA Graphs — iamfakhrealam · 2026-07-24
- MiniCPM-Robot Enables Fully Local Robot Inference: Offline Tracking at 33Hz — HeyToha · 2026-07-25
- New open embodied-AI stack emphasizes contextual memory and fully local RobotTrack inference — HeyToha · 2026-07-25