Three Paths to Embodied Intelligence: VLA, World Models, or 'Instinct'?
VLA、世界模型还是"本能"？具身智能的三条路线走到哪一步
In 2026, embodied intelligence stands at a critical juncture between technological exploration and industrial application. The article examines the current state of three technical approaches: VLA, world models, and 'instinct-driven' systems, noting that data-driven methods remain dominant. However, the 'instinct-driven' path, exemplified by the Oakwood Fruit robot, is advancing through zero-data cold start and tactile perception to achieve operational capabilities. The article also analyzes data challenges, commercialization paths, and industry expectations for a 'ChatGPT moment,' emphasizing the importance of real-world deployment and reliability.
Source: Chinese robotics — Robot learning research · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
Natus 运行在端侧，无需云端大模型推理，端侧频率达 250Hz、单次反馈闭环约 4 毫秒；它不预设任何动作轨迹，只定义 " 期望状态 "—— 抓取时力度刚好、不捏坏、不滑落 —— 机器人自主探索收敛到目标，同一任务可产生多条可行动作路径。
Open source S11
Source：Chinese robotics — Robot learning research · ithome.com