面壁智能 (Mianmian AI) has released the MiniCPM-Robot series, a suite of open-source models designed for embodied intelligence. The MiniCPM-RobotManip model, with only 1.5 billion parameters, demonstrates competitive performance in vision-language-action tasks, rivaling larger models in benchmarks and significantly reducing inference latency. Additionally, the MiniCPM-RobotTrack model, at 0.9 billion parameters, excels at robot tracking and navigation, outperforming existing open-source solutions in various challenging scenarios. These models, along with the PhyAI inference framework, aim to accelerate the development and deployment of intelligent robots capable of complex physical interactions. AI
IMPACT Accelerates development of embodied AI agents by providing efficient, open-source models for physical tasks and navigation.
RANK_REASON Open-source release of new embodied AI models by a notable AI lab. [lever_c_demoted from frontier_release: ic=2 ai=1.0]
Read on Hugging Face Daily Papers →
- MiniCPM-Robot
- MiniCPM-RobotManip
- MiniCPM-RobotTrack
- MiniCPM-V
- OpenTrackVLA
- TrackVLA++
- artificial general intelligence
- MiniCPM-V 4.6
- Physical intelligence
- Vision-language-action policies
- World Action Models
- World Models
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
