NVIDIA 发布 Alpamayo 2 Super：面向 Robotaxi 与自动驾驶的 34B 开源视觉-语言-动作模型

NVIDIA 发布了 Alpamayo 2 Super：https：//huggingface.co/nvidia/Alpamayo2-Super，一款 34B 参数的视觉语言动作（VLA）模式：https：//blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/l，采用开放商业许可，适用于自动驾驶。设计目标是长尾事件：传统检测与预测堆栈处理不佳的罕见多代理情形。该模型将基于NVIDIA Cosmos 3 Super Reasoner构建、经过强化学习后期训练的32B VLM骨干，与基于扩散的2.3B动作解码器配对。从一次全环绕摄像机视频的扫描中，它发出一个有计划的轨迹、对该轨迹的因果解释以及一个元动作。

是的，而且从第一天起就是商业用途。权重根据 OpenMDW-1.1 发布，这是 Linux 基金会对开放模型发行版的宽松许可;源代码为Apache 2.0。该许可证涵盖微调、衍生模型和商业再分发。NVIDIA 正在将 OpenMDW 应用于整个 Alpamayo 系列，因此早期为研发推出的版本现在可以在无需额外许可的情况下商业部署。

输入是多机位RGB视频、文本和带有时间戳的自我动态历史。经过验证的公共笔记本配置文件使用六个摄像头和四个历史帧。自我运动是三维平移加上一个3×3旋转矩阵，多时间步。

轨迹API返回64个航点，间隔0.1至6.4秒，间隔0.1秒。每个航点携带自我框架XYZ和一个3×3旋转矩阵。

训练数据大约是115,000小时的多机位驾驶视频，包含自我运动和轨迹标注。它包含约3,700,000条因果链（CoC）痕迹——结构化、因果关联的驱动决策解释。图像训练数据超过十亿张。

在LingoQA：https：//github.com/wayveai/LingoQA，Alpamayo 2 Super 在 Lingo-Judge 中得分为79.2，在近40个评估模型中排名第一。在NVIDIA的测试中，它领先Qwen2.5-VL 72B17.0个百分点，Gemini 2.5 Pro领先15.1个百分点，GPT-4o领先23.2个百分点。

规划工作还需要关注另外两个数字。在 PhysicalAI-AV-NuRec：https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles-NuRec 数据集的 910 个场景上，使用 AlpaSim：https://github.com/NVlabs/alpasim 进行闭环评估时，AlpaSim 得分为 1.50 ± 0.13。在 PhysicalAI-AV：https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles 数据集的 937 个具有挑战性的样本上进行开环评估时，6.4 秒时的 minADE₆ 为 0.911 米。

对于每个驾驶场景，模型会产生一条轨迹、一条说明决策的 CoC 轨迹、一项元动作（如让行或变道）、推理自动标签，以及带有二维定位的视觉问答。

这种组合使得该版本在操作上非常有趣。开发者可以将模型所观察的内容与其选择的动作联系起来。CoC 轨迹可以与 NVIDIA Halos：https://www.nvidia.com/en-us/ai-trust-center/halos/autonomous-vehicles/ 的安全验证工作流程集成，并支持符合 ISO/PAS 8800 标准的 AI 安全。

NVIDIA 表示，当模型作为专有车队数据的自动标注工具使用时，可将标注周期从数月缩短至数天。

需要与我们合作以推广您的 GitHub 仓库或 Hugging Face 页面或产品发布或网络研讨会等吗？请联系我们： https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq 是 Marktechpost Media Inc. 的首席执行官。作为一位有远见的企业家和工程师，Asif 致力于利用人工智能的潜力造福社会。他最近的项目是推出人工智能媒体平台 Marktechpost，该平台以对机器学习和深度学习新闻的深入报道而著称，技术上严谨且易于广泛受众理解。该平台每月浏览量超过 200 万次，显示了其在受众中的受欢迎程度。

[免费指南] 保护 AI 代理、MCP 服务器和大语言模型应用程序：https://pxllnk.co/lxn88m

NVIDIA has released Alpamayo 2 Super：https://huggingface.co/nvidia/Alpamayo2-Super , a 34B-parameter vision-language-action (VLA) mode：https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/l for autonomous driving, under an open commercial license. The stated design target is the long-tail events: rare, multi-agent situations that conventional detection-and-prediction stacks handle poorly. The model pairs a 32B VLM backbone, built on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, with a 2.3B diffusion-based action decoder. From one pass over full-surround camera video it emits a planned trajectory, a causal explanation of that trajectory, and a meta-action.

Yes, and for commercial use from day one. The weights are released under OpenMDW-1.1, the Linux Foundation’s permissive license for open model distributions; source code is Apache 2.0. The license covers fine-tuning, derivative models and commercial redistribution. NVIDIA is applying OpenMDW across the entire Alpamayo family, so earlier releases introduced for R&D are now deployable commercially without additional permission.

Inputs are multi-camera RGB video, text, and egomotion history with timestamps. The validated public notebook profiles use six cameras and four historical frames per camera. Egomotion is 3D translation plus a 3×3 rotation matrix, multi-timestep.

The trajectory API returns 64 waypoints spanning 0.1 to 6.4 seconds at 0.1-second intervals. Each waypoint carries ego-frame XYZ and a 3×3 rotation matrix.

Training data is roughly 115,000 hours of multi-camera driving video with egomotion and trajectory annotations. It includes about 3,700,000 Chain-of-Causation (CoC) traces — structured, causally linked explanations of driving decisions. Image training data exceeds one billion images.

On LingoQA：https://github.com/wayveai/LingoQA, Alpamayo 2 Super records a Lingo-Judge score of 79.2 and ranks first among nearly 40 models evaluated. In NVIDIA’s testing it beat Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1, and GPT-4o by 23.2.

Two more numbers matter for planning work. Closed-loop evaluation with AlpaSim：https://github.com/NVlabs/alpasim on 910 scenarios from the PhysicalAI-AV-NuRec：https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles-NuRec dataset gives an AlpaSim score of 1.50 ± 0.13. Open-loop evaluation on 937 challenging samples from the PhysicalAI-AV：https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles dataset gives minADE₆ at 6.4s of 0.911m.

For each driving situation, the model produces a trajectory, a CoC trace explaining the decision, a meta-action such as yield or lane change, reasoning auto-labels, and visual question answering with 2D grounding.

That combination is what makes the release interesting operationally. Developers can tie what the model observed to the action it chose. CoC traces integrate with NVIDIA Halos：https://www.nvidia.com/en-us/ai-trust-center/halos/autonomous-vehicles/ safety-validation workflows and support AI safety aligned with ISO/PAS 8800.

Used as an autolabeler on proprietary fleet data, NVIDIA says the model compresses annotation cycles from months to days.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.

[FREE GUIDE] Securing AI Agents, MCP Servers & LLM Apps ：https://pxllnk.co/lxn88m