ModelBest Open-Sources Embodied AI Dual Models

新智元 · wechat · 2026-07-19

At WAIC, ModelBest open-sourced the embodied AI series **MiniCPM-Robot**, which includes two models: - **MiniCPM-RobotManip**: A general VLA model primarily addressing the lack of contextual memory in robots. The article notes it scored **53.5** on RMBench, compared to just **10.4** for pi0.5, and ranks in the top tier on benchmarks like LIBERO. The core approach uses **extreme visual Token compression** to reduce the compute cost of inference with historical frames, making the cost of "memory-enabled" inference nearly identical to "memory-less" inference. It reports a single-step decision latency of **120ms** and supports long-horizon tasks and continuous operations like making a sandwich. - **MiniCPM-RobotTrack**: A tracking and navigation model natively compatible with the Unitree Go2Edu, supporting natural language following in local, offline environments. The article provides test results for STT/DT/AT and notes its stable operation in weak-network scenarios like real elevators and underground garages, with an end-to-end response time of roughly **180ms**. The article also introduces the accompanying inference framework **PhyAI**, claiming significant acceleration for various VLA/world models on hardware like RTX 5090 and H20. The overarching philosophy is to open-source the "model + inference framework + deployment solution" as a cohesive package to lower the barrier to embodied AI adoption.

Related event: ModelBest Open-Sources Embodied Model Series MiniCPM-Robot at WAIC(6 posts)→

Original post →

More from Embodied

Embodied channel →