具身 AI
MiniCPM-Robot Compresses Streaming Visual Memory and Action Models into 1.5B Parameters, Completes the Path to On-Device Go2 Deployment
OpenBMB has updated MiniCPM-Robot’s weights, code, and deployment documentation, enabling a single 1.5B VLA model to handle multiple manipulation tasks, while a separate 0.9B model performs target tracking. Streaming context and a custom inference engine reduce redundant visual computation, but the official benchmarks remain insufficient to demonstrate cross-robot generalization.
OpenBMB first released MiniCPM-Robot on July 19 and has continued updating its model cards and deployment materials over the past two days. The series includes MiniCPM-RobotManip and MiniCPM-RobotTrack. The former is a general-purpose 1.5B-parameter vision-language-action model that uses a single set of weights for both simulated and real-world manipulation. The latter is a 0.9B tracking policy designed for stationary, moving, and adversarial targets. The code and weights are licensed under Apache-2.0.
RobotManip’s key innovation is not merely reducing the parameter count, but avoiding the need to re-encode the full history during every control cycle. According to the project team, when retaining 60 frames, conventional recomputation requires approximately 125 TFLOPs per decision, whereas streaming inference reduces this to 3.3 TFLOPs. The number of visual tokens per frame is also compressed from 256 to 64. The model can maintain roughly one minute of visual memory. On an H100 using BF16 with a single-frame setting, model forward-pass latency is 120 milliseconds. However, this figure excludes autoregressive action decoding and therefore should not be interpreted as complete control latency.
The accompanying PhyAI inference engine uses CUDA Graphs and Triton fused kernels to increase inference frequency on an H20 from 10 Hz to a claimed 37 Hz. RobotTrack, meanwhile, provides a deployment workflow for the Unitree Go2 EDU using a Jetson Orin NX 16GB, TensorRT 10.7, and ROS 2 Humble. End-to-end latency is approximately 180 milliseconds, delivering more than 5 FPS. The default dry-run mode also reflects the fact that real-world control cannot be evaluated solely through offline scores.
Engineering teams should next focus on three questions: whether third parties can reproduce the reported LIBERO, RoboTwin2, and EVT-Bench results; whether streaming state drifts under occlusion, dropped frames, and prolonged operation; and how much redistillation or fine-tuning is required when transferring the same policy to robots with different cameras, joints, and control frequencies. At this stage, the work is better viewed as an experimental low-latency baseline than as a validated, hardware-agnostic general-purpose robot brain.