
ICRA 2026最佳论文奖,千寻智能首席科学家、清华高阳团队FP3入围
Quick Answer
The FP3 model, developed by Tsinghua University's Gao Yang and Spirit AI, is nominated for the ICRA 2026 Best Paper Award, showcasing a 1.3B parameter 3D strategy model that outperforms existing benchmarks by over 80% in unknown environments.
Quick Take
Spirit AI also announced a record 15 billion RMB funding and its model Spirit v1.6 topped the RoboArena benchmark, marking significant advancements in .
Key Points
- FP3 achieves over 90% task success rate in unknown environments with minimal training samples.
- Spirit AI's funding reaches 15 billion RMB, setting a new industry record.
- FP3 utilizes a 1.3B parameter diffusion Transformer architecture for enhanced performance.
- Spirit v1.6 model surpasses NVIDIA's Cosmos3 in RoboArena benchmarks.
- Chinese embodied AI startups are rapidly expanding, with several nearing IPO.
Source Excerpt
原文作者:谭梓馨原文链接:https://mp. weixin. qq. com/s/llcXE2be4oNWItL_0ydVZw2026年IEEE国际机器人与自动化会议(IEEE ICRA 2026)是机器人与自动化领域的顶级学术盛会,于6月1日至5日在奥地利维也纳举办。 今年,FP3、HITTER等多篇华人团队论文入围最佳论文奖提名,头部科技此前曾报道过HITTER,今天来看另一篇研究FP3。 FP3论文的导师之一高阳是清华大学跨学科信息科学研究院(IIIS)的助理教授,同时他也是国内具身独角兽千寻智能(Spirit AI)的联创兼首席科学家。 下面,一起来看看FP3做了哪些创新工作。 三维基座带来的改进基于海量多任务数据预训练的基座大模型在机器人领域展现出巨大潜力,但绝大多数机器人基础策略模型仅采用二维图像作为输入观测,缺失三维几何信息,而三维几何是机器人感知、理解真实三维空间的关键。 概括来说,FP3是一个面向机器人操作的大规模三维基座策略模型。 该模型参数大小为1.
3B,基于可规模化的扩散Transformer架构搭建,依托6万条含点云观测的运动轨迹完成预训练,凭借特有模型结构与多元化预训练数据,FP3可快速微调适配各类下游任务,泛化性能优异。 真机实测表明:仅需80组人工示教样本,FP3即可在包含全新未知物体的陌生场景中习得新任务,性能大幅领先现有机器人基座模型。 在自建的多项新任务数据集上验证FP3高效微调与强泛化特性,结果显示:仅使用单卡、两小时微调,相较主流基线,同场景性能平均提升60%,开放未知场景性能平均提升80%。 …
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from 雷峰网机器人
See more →
WAIC深度观察:具身数据,进入「开箱即用」时代
The WAIC forum highlights the urgent need for 'Model-Ready Data' in embodied intelligence, emphasizing high precision, multi-modality, and diversity as essential for effective physical interaction. Companies like JianZhi Robotics are pioneering solutions to bridge existing data gaps, ensuring that data is directly usable without further processing.

