SLAI Seminar 36th Session will be discussing the topic on "Open-Source Frameworks and Foundation Models Empowering Embodied Intelligence",  from 2:30pm-4pm, May 14 (Thursday) at Room B401, online participation is welcome.

 (Tencent Meeting ID: 884 315 040)

报告主题： 开源框架与基础模型赋能具身智能

时间：2026年5月14日（周四）下午2:30-4:00

地点: 深圳河套学院B401教室

线上参与：腾讯会议号884-315-040

主讲嘉宾：苏治中博士

主持人：姬艳丽教授

讲者简介About the Speaker:

苏治中博士现任地平线研究院副院长兼机器人实验室主任，全面负责地平线具身智能相关的技术研发工作。苏博士曾任地平线智能驾驶全栈算法高级总监，领导400余人团队完成地平线多款芯片上的多个场景的AI产品的研发，落地数十款车型，出货数百万套。苏博士在T-PAMI、TVCG、CVPR、ICCV、NeurIPS、ICRA等人工智能和机器人期刊和会议上发表论文并任评审；指导团队在ICDAR，COCO，Waymo，KITTI，nuScenes等国际AI竞赛或榜单拿到第一。加入地平线前，苏博士就职于百度，为深度学习实验室创始成员之一。

Dr. Zhizhong Su currently serves as the Vice President of Horizon Robotics Research Institute and Director of the Robotics Lab, where he is fully responsible for Horizon's technology research and development in embodied intelligence. Previously, Dr. Su was the Senior Director of Horizon's full-stack autonomous driving algorithm team, leading a team of over 400 members in the development of AI products across multiple scenarios on various Horizon chips. His work has been deployed in dozens of vehicle models and millions of units shipped. Dr. Su has published papers and served as a reviewer for prestigious journals and conferences in artificial intelligence and robotics, including T-PAMI, TVCG, CVPR, ICCV, NeurIPS, and ICRA. He has also led his team to top rankings in international AI competitions or benchmarks such as ICDAR, COCO, Waymo, KITTI, and nuScenes. Prior to joining Horizon, Dr. Su worked at Baidu as one of the founding members of the Deep Learning Lab.

报告摘要：

随着大模型技术的突破，具身智能正从实验室走向真实场景。本报告聚焦地平线的机器人实验室在具身智能领域的技术布局，核心围绕"数据—运控—导航—操作"全链条展开。

在数据上，我们自主研发了EmbodiedGen仿真数据引擎，融合Real2Sim数字孪生与AI生成双模式，一站式构建高保真可交互3D环境。运控领域，针对人形机器人全身运动控制难题，推出HoloMotion基础模型，具备实时、鲁棒、可泛化的运控能力，并配套完整的数据获取、强化学习多机并行训练与真机部署管线。导航领域，Memory、Think和Dream三个系列的工作使机器人在复杂室内环境中实现自主导航与目标定位。操作领域，HoloBrain基于多项创新技术实现操作技能的成功率和泛化提升。

本报告旨在分享地平线在具身智能开源生态建设中的实践与思考，探索开源框架与基础模型如何协同赋能，推动机器人从单一任务向通用智能演进。

Abstract:

With the breakthroughs in large model technologies, embodied intelligence is transitioning from laboratories to real-world scenarios. This report focuses on Horizon Robotics' technological roadmap in the field of embodied intelligence, centering on the entire chain of "data — motion control — navigation — manipulation."

In terms of data, we have independently developed the EmbodiedGen simulation data engine, which integrates Real2Sim digital twins and AI-generated dual modes to build high-fidelity, interactive 3D environments in an all-in-one manner. In motion control, addressing the challenges of whole-body control for humanoid robots, we have introduced the HoloMotion foundation model, which delivers real-time, robust, and generalizable motion control capabilities, along with a complete pipeline for data acquisition, multi-machine parallel reinforcement learning training, and real-world deployment. In navigation, our three series of works—Memory, Think, and Dream—enable robots to autonomously navigate and locate targets in complex indoor environments. In manipulation, HoloBrain leverages multiple innovative technologies to improve the success rate and generalization of manipulation skills.

This report aims to share Horizon Robotics' practices and reflections in building an open-source ecosystem for embodied intelligence, exploring how open-source frameworks and foundation models can work together synergistically to drive robots from single-task execution toward general-purpose intelligence.