苏度科技在2025年世界人工智能大会上展示了其基于物理智能的具身智能技术成果[1]。该公司展示了包括移动操作、双手协作、柔性操作、精密组装在内的10余个机器人技能[1]。
苏度科技采用了以大规模仿真数据为主、少量真机数据为辅的训练方式[1]。公司联合创始人韩铮指出，这一方法解决了机器人缺乏"互联网数据"导致的泛化难题[1]。该公司的首个机器人大模型预训练时几乎没有使用真机数据[1]。
从单个技能到10余个技能的突破仅用了3个月时间[1]，这背后是基础技能被复用、组合和迁移[1]。苏度科技致力于构建可复用的底层操作模型[1]。
成立仅一年，苏度科技估值已达到200亿元[1]。公司目前接近万卡级别算力投入[1]。为验证在制造业场景的落地能力，苏度科技已与头部电池企业宁德时代开展多工位合作，首个工位已完成概念验证[1]。
Sandu Technology unveiled its embodied intelligence technology based on physical intelligence at the 2025 World Artificial Intelligence Conference in Shanghai, demonstrating over a dozen skills including mobile manipulation and dual-arm collaboration [1]. The company, which was co-founded in 2025, has reached a valuation of 2 billion yuan just one year after its establishment [1].
The core innovation behind Sandu's approach lies in its training methodology, which relies primarily on large-scale simulation data supplemented by limited real-world robot data [1]. Co-founder Han Zheng explained that this strategy resolves the generalization challenge faced by robots due to their lack of "internet-scale data" [1]. The company has demonstrated remarkable efficiency in skill development, expanding from a single capability to over ten skills within three months, leveraging foundational skills that are reused, combined, and transferred across tasks [1].
Sandu is pursuing real-world validation of its technology through collaboration with CATL, a leading battery manufacturer, on multi-station manufacturing applications [1]. The first station has completed proof of concept testing [1]. The company is investing computational resources approaching ten thousand GPU-equivalents in scale and is focused on constructing reusable foundational operation models [1].
