IT之家 10 月 21 日消息,英伟达当地时间 10 月 21 日在 GTC 柏林大会上宣布扩展 Isaac 开放机器人开发平台,新增人形机器人参考平台、运行框架、仿真资产和数据集,并更新机器人学习工具,覆盖从训练数据生成、仿真训练到实体机器人部署的开发流程。
硬件维度,英伟达联合思灵机器人(Agile Robots)发布了人形机器人参考平台 Agile ONE SN,计划于 2027 年初推出。
该平台预规格英伟达 Isaac GR00T,并搭载 Jetson Thor 计算平台,主要面向机器人研究与开发。借助整合仿真、训练和部署工具,Agile Robots 表示,该平台可将相关环境的前期规格时间从数周缩短至数天。
英伟达创始人兼 CEO 黄仁勋表示,通用机器人工艺正在快速发展,机器人领域的“ChatGPT 时刻”正加速到来。英伟达称,此次更新旨在为开发者给出模型、数据、仿真工具和计算平台,协助他们将通用机器人工艺应用于实际产业情形。
思灵机器人 Agile ONE SN 使用 Agile ONE S 人形机器人本体,身高 1.4 米、重量 40 公斤,具有 31 个自由度,单臂负载为 3 公斤。其配备两只 WUJI Hand 2 五指灵巧手,每只手具有 20 个自由度,兼容 1kHz 力控制和触觉感知。需要注意的是,路透社报道该机器人重量约为 34 公斤,与英伟达官方公布的 40 公斤数据存在差异,现有信息尚不足以确认差异原因。
计算维度,Agile ONE SN 搭载英伟达 Jetson Thor,给出 2070 TFLOPS 的 FP4 算力和 128GB 统一内存。传感器涵盖头部及腕部摄像头、激光雷达和惯性测量单元,用于感知周围环境并兼容机器人与外部环境交互。
安全维度,该平台计划将机器人端安全协议与英伟达 Halos 全栈安全平台集成,当下这项集成仍待完成。英伟达表示,斯坦福大学、苏黎世联邦理工学院等研究机构已对 Agile ONE SN 表示兴趣。
路透社另报道称,Agile Robots 计划于 2027 年与英伟达合作生产最多 2000 台人形机器人,面向大学和企业研究实验室等客户。
软件维度,英伟达 Isaac Lab 3.0 已正式发布,作为开源机器人学习框架,兼容多 GPU 训练工作流,并可使用 Newton、PhysX 等物理引擎。该框架旨在协助开发者训练和优化机器人策略,覆盖从桌面工作站到数据中心的不同计算环境。
IT之家注意到,英伟达还推出开源运行框架 Isaac OS。该框架基于 Holoscan 构建,用于连接 AI 模型、传感器、控制平台和机器人硬件,兼容从仿真环境迁移至实体机器人,并给出确定性实时运行能力。
针对 DGX Station,英伟达同时发布了三套开源参考工作流。第一套借助强化学习优化 GR00T 1.7 模型;第二套利用 Isaac Lab 训练灵巧操作技能;第三套则借助 ChatGPT Astra 对 Cosmos 世界模型开展后训练,以生成合成训练数据。这些工作流旨在协助开发者整合机器人学习、仿真和数据生成环节。
数据维度,SimReady Central Library 给出超过 2000 项可用于仿真的机器人和工厂资产,涵盖多种机器人设备及工业情形资源。Human 1M 数据集则包含多种人类活动、同步三视角视频和三维手部追踪数据,可用于机器人学习与动作理解。两项资源均借助 Hugging Face 给出,其中完整版本的 Human 1M 数据集仅向 Cosmos Coalition 合作伙伴开放。
英伟达还介绍了多家使用其机器人开发工具的企业。ABB Robotics 将 Omniverse 库整合至 RobotStudio HyperReality,用于机器人平台安装前的仿真和虚拟调试;NEURA Robotics 则利用 Isaac Sim 和 Isaac Lab 生成合成训练数据,并在仿真环境中训练机器人技能。英伟达表示,KUKA、Humanoid、Sereact 和 Universal Robots 等企业也在使用其开放机器人平台相关工艺。
本文来源:https://www.ithome.com/1/016/128.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
On October 21, Nvidia announced the expansion of Isaac's open robot development platform at the GTC Berlin Conference, adding a humanoid robot reference platform, operating framework, simulation assets and datasets, and updating robot learning tools to cover the development process from training data generation, simulation training to physical robot deployment.
In the hardware dimension, Nvidia Agile Robots released Agile one SN, a reference platform for humanoid robots, which is scheduled to be launched in early 2027.
The platform is pre-specified as the Nvidia Isaac GR00T and is powered by the Jetson Thor computing platform, which is mainly for robotics research and development. With integrated simulation, training and deployment tools, Agile Robots says the platform can reduce the upfront specification time for relevant environments from weeks to days.
Huang Jenxun, founder and CEO of NVIDIA, said that the general robotics process is developing rapidly, and the "ChatGPT moment" in the robotics field is accelerating. Nvidia said the update aims to give developers models, data, simulation tools and computing platforms to help them apply general-purpose robotics processes to real-world industry situations.
The Agile one SN uses the Agile one S humanoid robot body, with a height of 1.4 meters, a weight of 40 kg, 31 degrees of freedom, and a one-arm load of 3 kg. It is equipped with two wuji Hand 2 five-finger dexterous hands, each with 20 degrees of freedom, compatible with 1kHz force control and tactile perception. It should be noted that Reuters reported that the robot weighs about 34 kg, which is different from the official 40 kg data released by Nvidia, and the available information is not enough to confirm the reason for the difference.
Computing dimension, Agile one SN is equipped with Nvidia Jetson Thor, giving 2070 TFLOPS of FP4 computing power and 128GB of unified memory. The sensors cover head and wrist cameras, lidar, and inertial measurement units, which are used to sense the surrounding environment and are compatible with the robot's interaction with the external environment.
Safety dimension, the platform plans to integrate the robot-side safety protocol with the Nvidia Halos full-stack safety platform, which is still to be completed. Nvidia said that research institutions such as Stanford University and ETH Zurich have expressed interest in Agile one SN.
Reuters also reported that Agile Robots plans to partner with Nvidia to produce up to 2,000 humanoid robots in 2027 for customers such as universities and corporate research labs.

Software dimension, Nvidia Isaac Lab 3.0 has been officially released as an open source robot learning framework, compatible with multi-GPU training workflows, and can use physics engines such as Newton and PhysX. The framework is designed to assist developers in training and optimizing robotics strategies to cover different computing environments from desktop workstations to data centers.
IT Home noticed that Nvidia also launched Isaac OS, an open source operating framework. Built on Holoscan, the framework connects AI models, sensors, console and robot hardware, is compatible with migration from simulation environments to physical robots, and gives deterministic real-time operational capabilities.
Nvidia also released three open source reference workflows for DGX Station. The first uses reinforcement learning to optimize the GR00T 1.7 model; the second uses Isaac Lab to train dexterity skills; and the third uses ChatGPT Astra to post-train the Cosmos world model to generate synthetic training data. These workflows are designed to assist developers in integrating robotics learning, simulation, and data generation.
Data dimension, SimReady Central Library gives more than 2000 robot and factory assets that can be used for simulation, covering a variety of robot equipment and industrial situation resources. The Human 1M dataset contains a variety of human activities, synchronized three-view video and three-dimensional hand tracking data, which can be used for robot learning and motion understanding. Both resources are presented with the help of Hugging Face, where the full version of the Human 1M dataset is only available to Cosmos Coalition partners.
Nvidia also introduced a number of businesses that use its robotics development tools. ABB Robotics integrates the Omniverse library into RobotStudio HyperReality for simulation and virtual debugging before the robot platform is installed; Neura Robotics generates synthetic training data using Isaac Sim and Isaac Lab and trains robotics skills in a simulation environment. Companies such as KUKA, Humanoid, Sereact and Universal Robots are also using their open robotics platform-related processes, Nvidia said.
This article is from: https://www.ithome.com/1/016/128.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.