ModelBest Open-Sources MiniCPM-Robot Embodied Models
面壁智能 · wechat · 2026-07-20
ModelBest, in collaboration with OpenBMB, has released and open-sourced its first embodied intelligence series, MiniCPM-Robot, which includes two models:
- MiniCPM-RobotManip: A 1.5B VLA model for robotic manipulation, trained based on MiniCPM-V4.6. It emphasizes small parameters, high performance, and visual Token compression. Officials claim it reaches the first tier on certain VLA benchmarks, and its in-context memory performance on RMBench significantly outperforms several mainstream models.
- MiniCPM-RobotTrack: A 0.9B visual instruction tracking model designed for mobile robot following/navigation, supporting local offline deployment. On real robots, it utilizes native cameras and local compute to perform single-object, dynamic multi-object, and fuzzy object tracking, stably achieving 5+Hz with an end-to-end latency of approximately 180ms.
The article also introduces its self-evolving data pipeline, DAgger-style data closed-loop, and the PhyAI inference framework tailored for Physical AI. The official direction is to deploy these embodied models and local execution capabilities in weak-network/disconnected scenarios such as inspection, companionship, and security.
Related event: Modelbest Open-Sources Embodied AI Model Series MiniCPM-Robot(6 posts)→
More from Embodied
- Inside Ant Group's play at WAIC-style expo: AI and hardware vendors settle into new division of labor — 智东西 · 2026-09-11
- Amazon and Google sold 600M+ smart speakers, so why no AGI-era successor? — julianlehr · 2026-09-11
- ECCV26 Oral: Flow Matching Enables Single-Stage Multi-View Point Cloud Registration — ducha_aiki · 2026-09-11
- Polish developers build iPhone app that detects nearby Meta smart glasses — Low-Honeydew6483 · 2026-09-11
- Ant's Afu health AI hits 150M users, unveils AI+hardware health alliance at Bund Summit — APPSO · 2026-09-11
- Johns Hopkins Launches Full-Stack Hands-on Robot Learning Class with SO-101 Arm Kits — _krishna_murthy · 2026-09-11