MiniCPM-Robot Enables Fully Local Robot Inference: Offline Tracking at 33Hz
HeyToha · x · 2026-07-25
The author points out that most current robot demos are fragile, relying heavily on stable Wi-Fi and cloud compute, which causes them to fail under weak networks or long tasks. The newly released MiniCPM-Robot series addresses this by combining contextual memory with fully local tracking.
The release includes three core components:
- MiniCPM-RobotTrack: Achieves fully offline target tracking on the Unitree Go2 Edu robot dog with 180ms end-to-end latency (5+ Hz). It continues making decisions even when physically disconnected from the internet.
- MiniCPM-RobotManip: A 1.5B VLA (Vision-Language-Action) model that utilizes historical visual context during multi-step tasks (e.g., remembering button presses, moved objects, or resuming after occlusion) without exploding inference costs.
- PhyAI Open-Source Framework: An inference infrastructure for Physical AI that uses CUDA Graph optimization and custom Triton kernels to boost inference rates from 10 Hz to 33 Hz or even 36 Hz.
Additionally, the model uses visual-token compression to cut visual processing computation by about 50% while remaining competitive with Qwen-VLA and π0.5, effectively reducing hardware costs and power draw.
Related event: OpenBMB Open-Sources MiniCPM-Robot for Offline Embodied AI(10 posts)→
More from Embodied
- Bedrock Robotics raises $600M as its fully autonomous excavators do paid work on live job sites — Sethwinterroth · 2026-10-27
- AW-0 hits 80.8% on RoboTwin tasks, beating strongest baseline by 16 points — trevordarrell · 2026-10-27
- AW-0's code stream fixes world model prediction: picks mango where baseline grabs apple — trevordarrell · 2026-10-27
- AW-0 runs at 230ms on a single RTX 5090, 2.4x faster than autoregressive code planning — trevordarrell · 2026-10-27
- AW-0 debuts: one generative world action model plans in code, predicts, and controls robots — trevordarrell · 2026-10-27
- Skyfall AI's SKY-001 world model turns sketches into CAD builds in 3.9s, vs 199s for GPT-6 Astra — rohanpaul_ai · 2026-10-27