Core Insights - The article discusses the launch of Depth Any Panoramas (DAP), a foundational model for panoramic depth estimation, which addresses the challenges of data scarcity and model generalization in spatial intelligence [1][19]. Data and Model Development - DAP is trained on an unprecedented scale of 2 million (2M) panoramic images, significantly surpassing previous datasets like Stanford2D3D and Matterport3D, which had only tens of thousands of images [6][7]. - The model utilizes a three-stage pseudo-labeling pipeline to refine the quality of depth estimation from unlabelled panoramic images, ultimately creating a robust training dataset [10][11]. Performance and Benchmarking - DAP has demonstrated superior performance in various benchmarks, achieving significant reductions in absolute relative error (AbsRel) and root mean square error (RMSE) across indoor and outdoor datasets [14][17]. - In zero-shot testing, DAP outperformed existing models, showcasing its strong generalization capabilities and effective depth prediction in complex environments [13][16]. Technological Innovations - The model incorporates advanced features such as a distance-adaptive range mask head, allowing it to adjust depth perception based on different application scenarios [16]. - DAP employs multi-dimensional geometric optimization techniques to ensure sharp edges and accurate geometric structures in depth maps, addressing common issues like depth holes and structural distortion [16]. Industry Implications - The introduction of DAP marks a significant milestone in panoramic depth estimation, enabling advancements in autonomous driving, robotics, and VR/AR content creation by providing a low-cost method for depth acquisition [19][20]. - The project has been made open-source, allowing broader access to its technology and fostering further innovation in the field of spatial intelligence [20].

点击下方 卡片 ，关注" 具身智能 之心 "公众号 编辑丨 具身智能之心 本文只做学术分享，如有侵权，联系删文 >> 点击进入→ 具身智能之心 技术交流群 更多干货，欢迎加入国内首个具身智能全栈学习社区 ： 具身智能之心知识星球 (戳我) ， 这里包含所有你想要的。 ★ 上次VLA模型+真机部署的圆桌受到了行业的一致好评。最近平台的同学也一直在整理对话的文稿，今天就为大家分享下第一部分" VLA的架构和模型 "相关内 容。 张强老师： 好，感谢主持人介绍，大家好，我是张强。我来自北京人形机器人中心，主要研究方向和研究背景都是在做人形机器人，大概从2021年开始做人形机器人。先后在 Fourier、GR-1 和 Embodied机器人，包括我们现在的天工平台上做了一些研究。我主要做的研究方向是运动控制，VLA 和一些基于人形机器人的世界模型和具身智 能大模型，希望大家关注我们的工作，然后今天也很高兴跟各位嘉宾。很高兴接受具身智能之心的邀请，很高兴跟各位嘉宾在一起讨论一下相关的问题，谢谢！ 完整内容欢迎加入我们的具身社区获取： 具身智能之心知识星球 主持人： 好，那我们就正式开始，那么欢迎大家来到具身智能之心的圆 ...

Core Insights - The robotics field is still in its early stages, as highlighted by Jim Fan, NVIDIA's robotics head, indicating a lack of standardized evaluation metrics and the disparity between hardware advancements and software reliability [1][8][11]. Group 1: Hardware and Software Disparity - Current advancements in robotics hardware, such as Optimus and e-Atlas, outpace software development, leading to underutilization of hardware capabilities [14][15]. - The need for extensive operational teams to manage robots is emphasized, as they do not self-repair and face frequent issues like overheating and motor failures [16][17]. - The reliability of hardware is crucial, as errors can lead to irreversible consequences, impacting the overall patience and scalability of the robotics field [18][19]. Group 2: Benchmarking Challenges - The lack of consensus on benchmarking in robotics is a significant issue, with no standardized hardware platforms or task definitions, leading to everyone claiming to achieve state-of-the-art (SOTA) results [20][21]. - The field must improve reproducibility and scientific standards to avoid treating them as secondary concerns [23]. Group 3: VLA Model Insights - The Vision-Language-Action (VLA) model is currently the dominant paradigm in robotics, but its reliance on pre-trained Vision-Language Models (VLM) presents challenges due to misalignment with physical world tasks [25][49]. - The VLA model's performance does not scale linearly with VLM parameters, as the pre-training objectives do not align with the requirements for physical interactions [26][51]. - Future VLA models should integrate physical-driven world models to enhance their ability to understand and interact with the physical environment [50]. Group 4: Data Importance - Data plays a critical role in shaping model capabilities, with the need for diverse data sources and collection methods being highlighted [31][43]. - The emergence of new hardware and data collection methods, such as Generalist and Egocentric-10K, demonstrates the growing importance of data in the robotics field [36][42]. - The current data collection strategies remain open-ended, with various approaches still being explored [43]. Group 5: Industry Trends - The robotics industry is projected to grow significantly, from $91 billion currently to $25 trillion by 2050, indicating a strong future potential [57]. - Major tech companies, excluding Microsoft and Anthropic, are increasingly investing in robotics software and hardware, reflecting the sector's attractiveness [59].

Core Viewpoint - The article discusses the advancements in the π series within the VLA (Vision-Language-Action) field, highlighting its role in transforming robotic learning paradigms and industry applications through continuous technological breakthroughs [2]. Group 1: Technological Advancements - The π0 model introduces Flow Matching for continuous action trajectory prediction, overcoming traditional discrete action precision limitations, providing a foundation for millimeter-level operations in precision manufacturing and autonomous driving scenarios [3]. - The π0.5 model features heterogeneous task collaborative training and hierarchical reasoning, achieving a 94% success rate in generalizing complex tasks in unfamiliar environments, while reducing data costs by 90% through human video training [3]. - The π0.6 model employs RECAP reinforcement learning for zero-shot generalization and efficient fine-tuning, surpassing human efficiency and precision in real-world applications, facilitating flexible production [3]. Group 2: Industry Impact - The π series models serve as a core reference for numerous VLA models in the industry since 2025, enabling the transition of general robots from laboratory settings to real-world applications in industrial manufacturing and home services [3]. - Companies are building their own demo machines based on the π series, such as for tasks like folding clothes and unpacking, indicating the practical implications of the technology [3]. Group 3: Learning and Development Challenges - Many beginners face difficulties in optimizing data and training VLA models based on the π series, with some spending up to six months without achieving satisfactory results [5]. - The article emphasizes the need for guided learning to help individuals gain practical experience and project work for job applications [6][11]. Group 4: Educational Offerings - The company offers a comprehensive course that covers hardware, data collection, VLA algorithms, evaluation, simulation, deployment of mainstream VLA models, and various real machine experiments [13][14]. - Participants in the course will receive a SO-100 robotic arm, enhancing hands-on learning opportunities [16]. Group 5: Target Audience - The course is aimed at individuals seeking practical experience in the VLA field, including students and professionals transitioning from traditional computer vision, robotics, or autonomous driving sectors [24].

点击下方 卡片 ，关注" 具身智能 之心 "公众号 作者丨 Younggyo Seo等 编辑丨具身智能之心 本文只做学术分享，如有侵权，联系删文 >> 点击进入→ 具身智能之心 技术交流群 更多干货，欢迎加入国内首个具身智能全栈学习社区 ： 具身智能之心知识星球 (戳我) ， 这里包含所有你想要的。 在人形机器人控制领域，强化学习（RL）虽已实现从仿真到现实的迁移，但高维动作空间、强域随机化需求导致训练周期冗长，严重制约迭代效率。 亚马逊 FAR 实验室团队提出的快速强化学习方案 ，以优化后的离线 RL 算法（FastSAC、FastTD3）为核心，通过 "算法调优 - 极简奖励设计 - 大规模并行仿真" 的 三位一体技术体系，首次实现单 GPU 15 分钟训练出鲁棒人形机器人 locomotion 政策，同时支持全身运动追踪任务的快速部署，彻底重构了人形机器人 sim-to-real 的迭代范式。 论文题目：Learning Sim-to-Real Humanoid Locomotion in 15 Minutes FastSAC-Humanoid — Project Page：https://youngg ...

点击下方 卡片 ，关注" 具身智能 之心 "公众号 机器人算法这么多？为什么很难走进真实场景？ 今年以来，围绕机器人操作任务，国内外团队产出了大量的工作。从physical intelligence 到国内的具身独角 兽、高校，不断在刷新各类指标，提升泛化性能。许多VLA和RL框架也逐渐完善，git上star动辄2k+，不少研 究团队持续在维护使用。 但也有一个很明显的问题，为什么VLA在真实场景中应用的很少？有些机器人在各类展会上的效果还算可 以，但稍微变动场景，就感觉"失明"一样，看着"张牙舞爪"。特别是叠衣服、拆箱子等日常生活中的一些任 务，动作固化、动作不够优雅是常态。 这个问题的根因是模型不够泛化，做过模仿学习的童鞋应该知道，模型如果"泛化性"不够，很难真的应用。 特别是具身机器人的开放场景，可能有N种动作，需要大量的数据喂给模型。 行业内一些数据采集成本高、周期长，难以规模化，定制化属性高。这就引出了一个非常重要的问题，如何 有效获取大规模高质量数据，让模型能够"泛化"，能够理解任务对应的行为和操作方式。 从成本和规模上已演变出四条数据路线 目前业界已形成多种具身数据获取方案，不同方案在数据质量、 ...

>> 点击进入→ 具身智能之心 技术交流群 更多干货，欢迎加入国内首个具身智能全栈学习社区 ： 具身智能之心知识星球 (戳我) ， 这里包含所有你想要的。 在机器人操纵与视觉 - 语言 - 动作（VLA）模型研究领域，人类凭借对历史信息的连贯理解与多步动作规划，能轻松完成复杂长时程任务（如制作披萨）。但现有 技术多聚焦于短时任务，在长时程场景中面临时序上下文缺失、状态漂移、资源消耗过大等挑战。 中科院、国科大与微软研究院联合团队提出的 LoLA 框架 ，以 "长时程潜动作学习" 为核心，通过 "多模态编码 - 状态感知融合 - 动作生成" 的三层架构，首次实现 了长时程语言引导机器人操纵的高效执行，为通用机器人在真实场景的落地提供了全新解决方案。 论文题目：LoLA: Long Horizon Latent Action Learning for General Robot Manipulation 核心亮点：长时程多模态融合编码、状态感知潜表征模块、跨平台泛化能力、仿真与真实场景双重验证 点击下方 卡片 ，关注" 具身智能 之心 "公众号 作者丨 Xiaofan Wang等 编辑丨具身智能之心 本文只做学术 ...

点击下方 卡片 ，关注" 具身智能 之心 "公众号 作者丨 Younggyo Seo等 编辑丨具身智能之心 本文只做学术分享，如有侵权，联系删文 >> 点击进入→ 具身智能之心 技术交流群 更多干货，欢迎加入国内首个具身智能全栈学习社区 ： 具身智能之心知识星球 (戳我) ， 这里包含所有你想要的。 在人形机器人控制领域，强化学习（RL）虽已实现从仿真到现实的迁移，但高维动作空间、强域随机化需求导致训练周期冗长，严重制约迭代效率。 亚马逊 FAR 实验室团队提出的快速强化学习方案 ，以优化后的离线 RL 算法（FastSAC、FastTD3）为核心，通过 "算法调优 - 极简奖励设计 - 大规模并行仿真" 的 三位一体技术体系，首次实现单 GPU 15 分钟训练出鲁棒人形机器人 locomotion 政策，同时支持全身运动追踪任务的快速部署，彻底重构了人形机器人 sim-to-real 的迭代范式。 论文题目：Learning Sim-to-Real Humanoid Locomotion in 15 Minutes FastSAC-Humanoid — Project Page：https://youngg ...

Core Viewpoint - The article discusses the advancements in the π series within the VLA (Vision-Language-Action) field, highlighting its role in transforming robotic learning paradigms and industry applications through continuous technological breakthroughs [2]. Summary by Sections Technological Advancements - The π series introduces significant innovations such as Flow Matching for continuous action trajectory prediction, which overcomes traditional discrete action precision limitations, providing a foundation for millimeter-level operations in precision manufacturing and autonomous driving by 2024 [3]. - By April 2025, π0.5 will implement heterogeneous task collaborative training and hierarchical reasoning, achieving a 94% success rate in generalizing complex tasks in unfamiliar environments, while reducing data costs by 90% through human video training, addressing industry data scarcity issues [3]. - The π0.6 model, expected by November 2025, will utilize RECAP reinforcement learning for zero-shot generalization and efficient fine-tuning, surpassing human efficiency and accuracy in real-world applications, achieving a 100% task completion rate and rapid reconfiguration within minutes, thus facilitating flexible production [3]. Industry Impact - The π series models are set to guide general robotics from laboratory settings to real-world applications in industrial manufacturing and home services, becoming a core reference for numerous VLA models in the industry since 2025 [3]. - Companies are building their own demo machines based on the π series, such as for tasks like folding clothes and unpacking, indicating a significant industry response to advancements in physical intelligence [3]. Learning and Development Challenges - Many learners face difficulties in effectively utilizing the π series for data and VLA model training, with some spending up to six months without achieving satisfactory results [5]. - There is a demand for guided learning experiences, as many students express a desire for mentorship to enhance their project portfolios for job applications [6][11]. Educational Initiatives - The company "具身智能之心" has replicated π0, π0.5, ACT, and GR00T methods to address the lack of real machines and project guidance for learners [7]. - A new course titled "VLA Small Class Course for Practical and Job-Oriented Learning" has been developed in collaboration with VLA experts to help students effectively learn and apply VLA technologies [8][13]. - The course covers a comprehensive curriculum including hardware, data collection, VLA algorithms, simulation, and various real machine experiments, aimed at providing a robust learning experience [13][14]. Course Details - The course includes a complimentary SO-100 robotic arm for participants, enhancing hands-on learning opportunities [17]. - It targets individuals seeking practical experience in the VLA field, including students and professionals transitioning from traditional fields [25]. - The course is set to commence on December 30, 2025, with a structured schedule for subsequent sessions [26][28].

点击下方 卡片 ，关注" 具身智能 之心 "公众号 作者丨 Jai Bardhan等 编辑丨具身智能之心 本文只做学术分享，如有侵权，联系删文 >> 点击进入→ 具身智能之心 技术交流群 更多干货，欢迎加入国内首个具身智能全栈学习社区 ： 具身智能之心知识星球 (戳我) ， 这里包含所有你想要的。 核心背景与问题 Vision-Language-Action（VLA）模型让机器人能够理解自然语言指令并执行操纵任务，但泛化能力评估一直是关键挑战——真实世界评估成本高、可重复性差，而 现有仿真基准存在明显缺陷：扰动类型有限、缺乏高保真视觉效果和真实的机器人控制对齐，导致仿真与真实世界性能脱节（即"现实-仿真差距"）。 为解决这一问题， 捷克理工大学,阿姆斯特丹大学的研究团队 构建了REALM：一个高保真仿真环境与基准，核心目标是建立仿真与真实世界性能的强相关性，实现 大规模、低成本的VLA模型泛化能力评估。其核心突破在于三点：高保真视觉与控制对齐的仿真环境、覆盖多维度扰动的系统评估方案、经实证验证的真实-仿真 性能关联性。 相关工作与差异化优势 现有机器人操纵泛化基准多依赖仿真，但存在显著局限：GemBench、 ...