OpenBMB's new robot release does not start by making the model bigger. It starts by asking whether a robot can keep acting when the network is weak and the onboard computer is small.
The MiniCPM-Robot family was made public on GitHub on July 19. The first two models split the work: MiniCPM-RobotManip is a 1.5B VLA model for manipulation, while MiniCPM-RobotTrack is a 0.9B model for target following and navigation. PhyAI, an inference framework for physical AI deployment, ships alongside them.
Latency is the first test
The official README reports 97.5 on LIBERO easy and 91.3/91.6 on RoboTwin2 clean and random for MiniCPM-RobotManip. On RMBench, which stresses memory over earlier observations, the model reaches 53.3, compared with 10.4 for π0.5. The model card also says visual tokens are compressed from 256 to 64 per frame, and H100 BF16 model-forward latency is about 120 ms per decision step versus 234 ms for π0.5.
Another official number is more revealing for long tasks: with 60 frames of history, traditional recomputation takes about 125 TFLOPs per step, while streaming inference takes 3.3 TFLOPs. That is roughly a 97% cut in per-step compute. For a robot arm, the gap shows up as smoother motion, not just a better benchmark row.
The dog has to follow locally
MiniCPM-RobotTrack targets a more physical use case: a Unitree Go2 Edu following a person from language and vision. Reported tracking rates are 89.8%, 73.4% and 80.4% across single-target, distractor-heavy and ambiguous-target tasks. The deployment stack is also explicit: Go2 Edu, Jetson Orin NX 16GB, Jetson Linux R36.5, CUDA 12.6, TensorRT 10.7 and ROS 2 Humble. The end-to-end local loop runs at 5+ FPS with roughly 180 ms latency.
That distinction matters. Cloud model latency and local motion-control latency are not the same metric. If the robot enters an elevator or underground car park, the camera, encoder, model, waypoint output and command path all have to stay inside the local loop.
Open source is useful only if it runs
The release includes weights, inference examples, Go2 deployment notes, environment instructions and safety defaults; the Hugging Face cards list Apache-2.0 licensing. PhyAI adds the engineering layer: QbitAI reported speedups of 2.85x, 1.82x and 2.28x on π0, π0.5 and GR00T on RTX 5090, and MiniCPM-Robot throughput moving from 10.12Hz to 33.28Hz with CUDA Graph, then to 36.77Hz on NVIDIA H20 with fused kernels.
Yao Yuan said many robot demos still tune heavily around one task, while MiniCPM's team wants to polish data, infrastructure and engineering paths before expanding to more scenarios.
The remaining test is not a launch demo. It is whether outside teams can reproduce the numbers, port the models beyond Go2, and publish long-running results in inspection, guiding or warehouse tasks.
Sources: QbitAI; OpenBMB GitHub README for release status, Apache-2.0 licensing, LIBERO/Calvin/RoboTwin/RMBench and EVT-Bench metrics; Hugging Face model cards for Manip/Track weights and local inference notes, CocoLoop, OpenBMB organization page for model update status.