Four Reported Routes for Embodied-AI Brain Architectures
A Lei Technology report republished by Cyzone groups embodied-AI brain development into four routes: world-model-native systems, VLA–world-model hybrids, brain-inspired architectures and unified models. It highlights ZhiSquare's three-layer NeuroVLA, assigning semantic understanding and task planning to the cortex, high-frequency motion coordination to the cerebellum and local motion execution and safety reflexes to the spinal cord. The report also cites AstraBrain and says Spirit v1.6 ranked first overall in RoboArena. It cautions that the four categories are an analytical grouping rather than a settled standard and may overlap.
The report says these architectures divide work among semantic planning, world simulation, motion coordination and real-time control. It also identifies power consumption, latency and data efficiency as central engineering challenges, while noting that no route has been established as the final solution.
编者按:本文来自微信公众号 雷科技(ID:leitech),编辑:罗超,创业邦经授权发布。
有句话是这么说的:人类与机器人的最大区别,是人类会预判。比如运动员,电竞选手等,有时候对手招还没出完,身体已经先动了。
但这个说法放在2026年的今天并不准确,如今,“预判”并不是人类和机器人之间绝对的界限,准确来说,机器人正在被技术赋予这种能力,与人类的区别在于,人类是基于意识、经验和直觉的自然预判,机器人则是基于数据、模型和算力的“计算预判”,也就是所谓的机器人大脑。
最近几个月,围绕这种让机器人“先想一步”的新闻特别多。
六月中的北京智源大会上,智平方发布了全球首个类脑式具身智能系统NeuroVLA,把机器人的大脑按人脑的皮层、小脑、脊髓三层结构重新设计了一遍。
同样是6月,千寻智能的Spirit v1.6在RoboArena评测里综合得分排到全球第一,超过英伟达的Cosmos3和Physical Intelligence的π0.5,成为首个登顶这个榜单的中国模型。
#NeuroVLA#Spirit v1.6#AstraBrain#EFM-1#WALL#GE 2.0#Gemini Robotics 2#Nvidia#Physical Intelligence#Google#OpenAI#智平方