点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 随着神经场景表征的发展，之前出现了一些方法尝试用神经辐射场重建街道场景，像Block-NeRF 。但是它无法处理街道上的动态车辆，而这是自动驾驶环境仿真 中的关键要素。最近一些方法提出将动态驾驶场景表示为由前景移动汽车和静态背景组成的组合神经表示。为了处理动态移动的目标车辆，这些方法利用跟踪的 车辆姿态来建立观察空间和规范空间之间的映射，在那里他们使用 NeRF 网络来模拟汽车的几何形状和外观。虽然这些方法产生了合理的结果，但它们仍然局限 于高训练成本和低渲染速度。基于这些前述工作，浙大提出了S treet Gaussians。笔者有幸参与了公司新一代闭环仿真系统的开发，花了几个月的时间，终于把基 于Street Gaussians的算法落地。今天就分享下自己的一些看法~ 下图是在Waymo数据集上的渲染结果。street gaussians的方法在训练半小时内以 135 FPS的速度产生高质量的分辨率为1066×1600渲染视角。这两个基于NeRF的方 法存在训练和渲染成本高的问题。 以前的方法通常面临训练 ...

以下文章来源于深蓝AI ，作者迟浩瀚 深蓝AI . 专注于人工智能、机器人与自动驾驶的学习平台。 作者 | 迟浩瀚 来源 | 深蓝AI 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 >>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 本文只做学术分享，如有侵权，联系删文 推荐语：自动驾驶领域的视觉-语言-动作（VLA）模型虽前景可期，却在非结构化边界场景中表现欠佳，主要归因于针对性基准数据的稀缺。为此，这 篇研究推出Impromptu VLA解决方案。核心贡献在于构建了Impromptu VLA数据集：该数据集包含8万余条精细构建的视频片段，提炼萃取自8个开源 大规模数据集中的200余万条原始素材。 本文由论文一作迟浩瀚 论 文题目： Impromptu VLA: Open Weights and Open Data for Driving Vision-Language-Action Models 论文作者： Haohan Chi, Huan-ang Gao, Ziming Liu, Jianing Liu, Chenyu Liu, Jinwei Li ...

点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 简单来说，占用网络的目的将空间划分成小网格，预测每个网格的占用情况，解决异形障碍物。 占用网络最重要的的功能之一就是建模异形障碍物，像倒地的树木枝干、不规则车辆等等，同时可以建模路面等其他背景元素。 通用障碍物自动标注怎么做？ 小林是一名主机厂的云端模型算法工程师，工作已经四五年了。这两天接到了车端报的Case，恶劣天气倒在地上的树干漏检，行车时紧急刹停差点酿成事故。。。不 出意外，原因最后排查到数据的问题，要求数据团队紧急补充训练数据。 小林为难的挠了挠头，这种异常case也太难解决了。检测没办法解决这种异常的占用问题，标注数据也从未见过，看来只能靠OCC来做下兜底了。想到这里，小林觉得 需要精标一小批数据提供车端先使用，再配合挖掘大模型和云端模型的自动标注模型快速迭代数据量才能保证车端模型的泛化，看来又要有一段时间的苦日子 了。。。 自从2022年特斯拉宣布Occupancy Network上车以来，当下占用网络已经作为各家纯视觉智驾方案的标配。目前OCC作为行车和泊车中的重要感知模块，对训练数据的 标注 ...

Core Viewpoint - WeRide has launched the HPC 3.0 high-performance computing platform, marking a significant advancement in autonomous driving technology and enabling the world's first mass-produced Level 4 autonomous vehicle, the Robotaxi GXR, powered by NVIDIA's DRIVE AGX Thor chips [1][4][10] Group 1: Product Development - The HPC 3.0 platform, developed in collaboration with Lenovo, features dual NVIDIA DRIVE AGX Thor chips and delivers up to 2,000 TOPS of AI compute, making it the most powerful computing platform for Level 4 autonomy [2][4] - The new platform reduces autonomous driving suite costs by 50% and cuts mass production costs to a quarter of its predecessor, HPC 2.0 [4][6] - HPC 3.0 consolidates key modules, which lowers the total cost of ownership (TCO) by 84% over its lifecycle compared to HPC 2.0 [4] Group 2: Safety and Compliance - HPC 3.0 is certified to AEC-Q100, ISO 26262, and IATF 16949 standards, with a failure rate below 50 FIT and a mean time between failures (MTBF) of 120,000 to 180,000 hours [5] - The platform is designed for 10 years or 300,000 km of use and can operate in extreme temperatures from -40°C to 85°C, meeting global VOCs environmental standards [5] Group 3: Strategic Partnerships - The collaboration with Lenovo and NVIDIA is highlighted as a major breakthrough in computing power and cost efficiency, enhancing vehicle reliability and responsiveness while significantly reducing deployment costs [6][7] - NVIDIA has been a strategic investor in WeRide since 2017, supporting the commercialization of autonomous driving solutions globally [8][9] Group 4: Market Position - WeRide is recognized as the world's first publicly listed Robotaxi company, having operated Robotaxis on public roads for over 2,000 days and tested its technology in over 30 cities across 10 countries [10][11] - The company has received autonomous driving permits in five markets: China, the UAE, Singapore, France, and the US, positioning itself as a leader in the autonomous driving industry [11]

点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 Orbis：突破驾驶世界模型长时域预测的技术挑战 弗莱堡大学提出基于流匹配的连续自回归世界模型 Orbis，在仅使用 280 小时单目视频数据和 4.69 亿参数 条件下，长时程预测性能超越现有方法——nuPlan 数据集上 FVD 降至 132.25（6秒预测），转弯场景轨迹 精度（Frechet Precision）提升至 0.47（Table 3）。 主要贡献： 算法框架： Vista 2.000 GEM Cosmos Orbis-MG (ours) 1.500 Orbis (ours) 1,000 500 12 0 8 4 Starting time (s) Figure 4: Video quality (FVD) over consecutive 4s time windows on nuPlan-turns. The x axis shows the starting time of the evaluated time window. 可视化 : Estimated trajectories ...

Core Viewpoint - Lucid Motors is considering a 1-for-10 reverse stock split to potentially enhance its stock attractiveness to institutional investors, despite the market generally viewing such moves negatively [2][9]. Group 1: Reverse Stock Split Details - Lucid Motors filed a preliminary proxy statement with the SEC for a special stockholders' meeting to authorize a reverse stock split at a ratio of 1-for-10 [2]. - A 1-for-10 reverse stock split means that Lucid will reduce its outstanding shares by a factor of 10, combining 10 old shares into one new share, which will increase the stock price proportionally [3]. - The reverse stock split does not change the overall market capitalization or the value of what investors own, as the stock price adjusts in accordance with the reduction in shares [4]. Group 2: Implications and Market Perception - While reverse stock splits are often associated with companies at risk of delisting due to low stock prices, Lucid's current trading price of approximately $3.15 per share does not place it in immediate danger [5][6]. - The potential upside of the reverse stock split is to attract institutional investors who may have policies against owning stocks below a certain price, thereby increasing demand for Lucid's shares [8]. - The market typically views reverse stock splits negatively, associating them with financial distress and falling stock prices, which may not be favorable for investment perception [9]. Group 3: Company Performance and Future Outlook - Lucid Motors is currently experiencing cash burn but has shown consistent delivery growth over the past seven quarters, with future performance hinging on the success of its new electric Gravity SUV and upcoming midsize platform [10]. - The company has sufficient liquidity to fund operations through the second half of 2026, indicating a level of financial stability despite the high-risk nature of its stock [11].

Alphabet旗下无人驾驶技术公司Waymo当地时间7月17日宣布，当日起，其无人驾驶服务在美国得州奥 斯汀覆盖更多地区。 ...

Wall Street expects a year-over-year increase in earnings on higher revenues when Mobileye Global (MBLY) reports results for the quarter ended June 2025. While this widely-known consensus outlook is important in gauging the company's earnings picture, a powerful factor that could impact its near-term stock price is how the actual results compare to these estimates.The earnings report, which is expected to be released on July 24, might help the stock move higher if these key numbers are better than expectati ...

Core Insights - WeRide has launched fully driverless Robobus operations at Resorts World Sentosa, Singapore, marking the first autonomous vehicle in Southeast Asia to operate without a safety officer on board [1][2][4] Company Developments - WeRide received approval from the Land Transport Authority of Singapore (LTA) after extensive testing and safety assessments, allowing the Robobus to offer fully autonomous rides to the public [2][9] - The Robobus has been operational since June 2024, transporting tens of thousands of passengers and maintaining a zero-incident safety record [2][5] - The Robobus operates on a fixed 12-minute loop connecting key points within Resorts World Sentosa, utilizing advanced LIDAR, cameras, and sensors for obstacle detection [5] Industry Impact - The launch is seen as a significant milestone for the future of mobility in Southeast Asia, with WeRide's vehicles expected to transform public transportation [4][6] - Singapore's government plans to integrate autonomous vehicles into the national public transport network by the end of 2025, aligning with WeRide's operations [6][9] - WeRide has established a dedicated R&D center in Singapore to advance autonomous vehicle innovation, supported by the Singapore Economic Development Board [6][12] Future Collaborations - WeRide aims to strengthen collaborations with LTA and various stakeholders to introduce more validated products and scalable business models across Singapore and Southeast Asia [11][12]

Group 1: Core Insights - Baidu's stock surged nearly 9% following a collaboration with Uber to introduce autonomous vehicles on Uber's platform outside of China and the U.S. [2] - The partnership marks a significant step in Baidu's efforts to internationalize its autonomous driving initiatives, with initial launches expected in Asia and the Middle East by the end of 2025 [3] - Baidu's Apollo Go division reported over 1.4 million rides in Q1 2025, a 75% increase year-over-year, and operates over 1,000 fully driverless vehicles across 15 cities [4] Group 2: Market Potential - Uber's human-driven rides generated a $375 billion annual revenue pool, indicating a substantial opportunity for the autonomous sector, which could potentially double the existing ride-hailing market to a $750 billion opportunity [4] - The demand for autonomous ride-hailing is expected to increase as users experience the benefits, with Robotaxis like Waymo showing higher customer retention and fewer accidents [4] Group 3: Challenges and Valuation - Baidu's stock has faced challenges due to a slower-than-expected post-Covid economic recovery in China, leading to reduced advertising revenue in its core search business [5] - The emergence of generative AI has created uncertainty for traditional search models, with competition from other Chinese tech giants like Alibaba and Tencent [5] - Baidu is currently valued at around $90 per share, trading at approximately 10x projected 2025 earnings, significantly lower than its nearly 40x multiple during the pandemic, with nearly $22 billion in net cash [6]