面向有强化学习基础的 Machine Learning PhD：从 world model、embodied imitation 到 VLA、action representation 与最新 VLA+world model 融合趋势。
This series is written for readers with reinforcement learning and machine learning background who want a structured path into embodied policies, VLA models, and the emerging VLA + world model frontier.
An RFM Implementation Blueprint: the Model Factory and the Data Flywheel
What we would build together You have already done the harder half: the robot is in production, deployed in real venues, running with VR teleoperation [blog: carnewschina 2026-04-13]. The other half is a robot foundation model and the data system that feeds it. What follows is that half — what gets built, in what order, and what evidence closes each step [arXiv:2511.19647]. The positioning is fixed from day one: what we are building is an on-device robotic foundation model. Both halves of that phrase are meant literally ...