GPT-6 Astra reshapes the embodied-AI debate
TMTPost reports that GPT-6 Astra can control robotic arms using visual inputs and task instructions, prompting debate over whether general multimodal models will displace specialized embodied models. Tests attributed to NVIDIA and Galaxy General reportedly found that Astra outperformed π0.5 on selected tasks, but also showed substantial latency and token costs. The article says demand for embodied data increased after the release, while practitioners remain divided over long-horizon reliability and the division of work between general and specialized models.
Astra's reported ability to handle robotic planning could intensify competition between foundation-model companies and embodied-AI startups, but its latency, cost, and long-task limitations leave specialized models relevant for motion control and real-world deployment.
(本文作者为 定焦One,钛媒体经授权发布)
文 | 定焦One(dingjiaoone),作者 | 王汉星,编辑 | 魏佳
一家成立一年半左右的具身智能数据服务公司,最近两周接到的数据采购订单超过了此前两个月的量,这在整个具身行业融资环境收紧的背景下就像是一针强心剂。
“订单不仅来自具身行业,一些曾经只询价不下单的基模大厂也开始下单了。”这家公司的创始人告诉「定焦One」。
变化始于GPT-6 Astra的发布,9月初,OpenAI发布最新一代模型后,社交媒体上出现大量用Astra控制机械臂做各种任务的测试,不少具身从业者发现,通用的多模态基础模型也具备了空间理解的能力,可以直接控制机器人执行任务,成功率甚至超过了一部分具身模型。
测试视频越传越广,“通用大模型迟早会‘吃掉’具身模型”“大语言模型将吞噬具身模型、具身智能只是通用大模型的一个子集”等声音开始在行业内流传。宇树科技创始人王兴兴甚至直言,明年GPT-7出来后,可能直接把具身智能问题打穿。但行业里也有不同的声音,有人认为当前的通用大模型还远没有到降维打击具身智能的地步,在长程任务成功率、延迟、成本等许多环节还存在明显的短板。
#GPT-6 Astra#OpenAI#NVIDIA#Physical Intelligence#π0.5#Galaxy General#Unitree Robotics#OpenAI Robotics#DeepSeek#Anthropic#Kimi#Lingyu Intelligence