英伟达2025 CES炸裂发布：自动驾驶与通用机器人ChatGPT时刻来临！｜新闻特写20250107

00:17:34
https://www.youtube.com/watch?v=4rz8b_UTO1s

摘要

TLDR视频中讨论了物理人工智能的未来愿景，主要是通过将AI应用在物理环境中，使其能够生成实际动作而不是文本。Nvidia 宣布了 Cosmos 世界基础模型，这是一种旨在理解物理世界的 AI 模型，并已开放许可进行使用。通过将 Cosmos 与 Omniverse 联合，提供了一个物理基础的多元宇宙生成器，可用于训练和模拟机器人在真实环境中的操作。这些工具的应用范围广泛，包括工业机器人、自主车辆等领域。Nvidia 还介绍了三台关键计算机系统：用于训练 AI 的 DGX 计算机、用于模拟和生成合成数据的 Omniverse 系统，以及用于部署 AI 的 AGX 计算机。此外，Nvidia 还推介了 Thor 处理器，这是一种适用于自动驾驶和其他机器人应用的强大处理器。

心得

  • 🔄 物理AI可将AI应用于实际环境执行动作，而非仅生成文本。
  • 🌍 Nvidia宣布了Cosmos，一个物理世界基础AI模型。
  • 🔗 Cosmos与Omniverse结合实现物理世界的模拟与AI生成。
  • 🚗 自动驾驶领域将借助这些技术进行合成数据生成和AI训练。
  • 💻 需要三种计算机系统：训练、模拟以及部署。
  • 🚀 Thor处理器，具备高处理能力，适合复杂机器人应用。
  • 🔍 AI在现实环境中的训练可通过数字孪生技术增强。
  • 🛡 Nvidia致力于确保AI系统的功能安全。
  • 🌐 这些技术涵盖了广泛的应用领域，如工业机器人和自动驾驶。
  • 📈 Nvidia的技术进步推动了物理AI和机器人行业的发展。

时间轴

  • 00:00:00 - 00:05:00

    在本视频的开头，讲者介绍了一种称为“物理AI”的概念，类似于大语言模型，但应用于物理世界的模型。这种模型不是通过生成文本，而是通过生成行动来响应物理指令。Nvidia推出的新产品“Cosmos”就是一个世界基础模型，旨在理解和模拟物理世界的功能。同时，该模型与Omniverse的结合使生成的内容有真实性和准确性，为机器人技术和工业AI提供新的科技基础。每个机器人系统需具备三台电脑：训练AI的电脑，部署AI的电脑，以及数字孪生模拟电脑。

  • 00:05:00 - 00:10:00

    接下来，视频讨论了Nvidia在工业领域三台电脑系统的策略：用于训练AI的DGX电脑，负责实际应用的AGX电脑，以及连接两者的数字孪生模拟系统。有诸多汽车公司已经开始与Nvidia合作开发新一代自主车辆，如丰田、特斯拉、沃尔沃等。Nvidia推出新一代汽车处理器“Thor”，相较于其前代产品更加高效，并且具有AI功能安全认证。这些技术进步展示了AI和自动驾驶车辆在工业中的广阔应用前景。

  • 00:10:00 - 00:17:34

    视频的最后部分展示了如何使用Omniverse和Cosmos来生成用于训练AI的合成数据。这种生成方法允许从现实驾驶场景创建大规模的数据，以支持自动驾驶车辆的发展。此外，还讨论了未来“类通用机器人时代”的来临，以及通过模拟生成训练数据以培养不同类型的机器人能力的可能性。Nvidia在推动机器人和自主车辆技术发展的同时，致力于构建大型世界基础模型，以更好地支持全球工业的转型。

思维导图

视频问答

  • 物理AI是什么？

    物理AI是指通过将AI应用于物理世界的模型，以产生实际动作，而不是生成文本。

  • 视频中提到的Cosmos与Omniverse是什么关系？

    Cosmos是一个世界基础模型，Omniverse是一个物理基础的模拟系统，二者结合可提供物理模拟和多元宇宙生成器。

  • Nvidia在视频中宣布了什么新项目？

    Nvidia宣布了开源的Cosmos世界基础模型，并展示了与Omniverse的结合。

  • 为何视频中提到的AI需要三个不同的计算机系统？

    AI的发展需要训练系统、数字孪生模拟系统和部署系统三者协同工作。

  • Cosmos模型为何重要？

    Cosmos模型可以帮助机器人理解和处理物理动态，进而提升工业AI和机器人技术。

  • 为什么与Omniverse结合是重要的？

    Omniverse的物理基础特征提供了AI生成内容的真实感和可靠性。

  • 视频中展示了什么应用案例？

    视频展示了自动化车辆如何利用AI和数字孪生技术生成合成数据来增强训练。

  • Nvidia的Thor处理器有什么特点？

    Thor是一种具有高处理能力的通用机器人计算机，适合自动驾驶和其他机器人应用。

  • 如何使用Cosmos和Omniverse训练AI模型？

    通过合成驾驶场景生成大量虚拟数据，借助真实世界数据校准和反馈进行模型训练。

  • Nvidia的新技术在自动驾驶领域的发展前景如何？

    Nvidia的技术将推动自动驾驶技术的发展，产生安全性和效率提升，成为一个大型产业。

查看更多视频摘要

即时访问由人工智能支持的免费 YouTube 视频摘要！
字幕
en
自动滚动:
  • 00:00:00
    okay let's talk about physical AI So
  • 00:00:02
    Physical
  • 00:00:03
    AI imagine
  • 00:00:07
    imagine whereas your large language
  • 00:00:09
    model you give it your context your
  • 00:00:14
    prompt on the left and it generates
  • 00:00:19
    tokens one at a time to produce the
  • 00:00:22
    output that's basically how it works the
  • 00:00:25
    amazing thing is this model in the
  • 00:00:27
    middle is quite large has billions of
  • 00:00:29
    parameters
  • 00:00:30
    the context length is incredibly large
  • 00:00:33
    because you might decide to load in a
  • 00:00:35
    PDF in my case I might load in several
  • 00:00:37
    PDFs before I ask it a question those
  • 00:00:41
    PDFs are turned into tokens the
  • 00:00:43
    attention the basic attention
  • 00:00:45
    characteristic of a transformer has
  • 00:00:47
    every single token find its relationship
  • 00:00:49
    and relevance against every other token
  • 00:00:53
    so you could have hundreds of thousands
  • 00:00:55
    of tokens and the computational load
  • 00:00:58
    increases quadratically
  • 00:01:00
    and it does this all of the parameters
  • 00:01:03
    all of the input sequence process it
  • 00:01:05
    through every single layer of the
  • 00:01:06
    Transformer and it produces one token
  • 00:01:09
    that's the reason why we need a
  • 00:01:10
    Blackwell and then the next token is
  • 00:01:13
    produced when the current token is done
  • 00:01:16
    it puts the current token into the input
  • 00:01:18
    sequence and takes that whole thing and
  • 00:01:21
    generates the next token it does it one
  • 00:01:22
    at a time this is the Transformer model
  • 00:01:26
    it's the reason why it is so so
  • 00:01:28
    incredibly effective computationally
  • 00:01:31
    demanding What If instead of PDFs it's
  • 00:01:35
    your surrounding and what if instead of
  • 00:01:37
    the prompt a question it's a request go
  • 00:01:40
    over there and pick up that you know
  • 00:01:42
    that box and bring it back and instead
  • 00:01:44
    of what is produced in tokens that's
  • 00:01:46
    text it produces action
  • 00:01:49
    tokens well that I just described is a
  • 00:01:54
    very sensible thing for the future of
  • 00:01:56
    Robotics and the technology is right
  • 00:01:58
    around the corner but what we need need
  • 00:02:00
    to do is we need to create the effective
  • 00:02:03
    effectively the world
  • 00:02:05
    model of you know as opposed to GPT
  • 00:02:09
    which is a language model and this world
  • 00:02:11
    model has to understand the language of
  • 00:02:13
    the world it has to understand physical
  • 00:02:15
    Dynamics we know that most models today
  • 00:02:19
    have a very hard time with and so we
  • 00:02:21
    would like to create a world we need a
  • 00:02:24
    world Foundation model today we're
  • 00:02:25
    announcing a very big thing we're
  • 00:02:28
    announcing Nvidia Cosmos a world
  • 00:02:32
    Foundation model that is designed that
  • 00:02:35
    was created to understand the physical
  • 00:02:37
    world and the only way for you to really
  • 00:02:39
    understand this is to see it today we're
  • 00:02:41
    announcing that Cosmos is open licensed
  • 00:02:45
    it's open available on
  • 00:02:53
    GitHub we hope we hope that this moment
  • 00:02:57
    and there's a there's a small medium
  • 00:02:58
    large for uh uh very fast models um you
  • 00:03:02
    know mainstream models and also teacher
  • 00:03:04
    models basically not knowledge transfer
  • 00:03:07
    models Cosmo Cosmos World Foundation
  • 00:03:10
    model being open we really hope will do
  • 00:03:13
    for the world of Robotics and Industrial
  • 00:03:15
    AI what llama 3 has done for Enterprise
  • 00:03:18
    AI the magic happens when you connect
  • 00:03:23
    Cosmos to Omniverse and the reason
  • 00:03:25
    fundamentally is this Omniverse is a
  • 00:03:30
    physics grounded not physically grounded
  • 00:03:33
    but physics grounded it's algorithmic
  • 00:03:36
    physics principled physics simulation
  • 00:03:39
    grounded system it's a simulator when
  • 00:03:42
    you connect that to
  • 00:03:44
    Cosmos it provides the grounding the
  • 00:03:47
    ground truth that can control and to
  • 00:03:50
    condition the Osmos generation as a
  • 00:03:53
    result what comes out of Osmos is
  • 00:03:55
    grounded on Truth this is exactly the
  • 00:03:57
    same idea as connecting a large language
  • 00:03:59
    model to a rag to a retrieval augmented
  • 00:04:03
    generation system you want to ground the
  • 00:04:05
    AI generation on ground truth and so the
  • 00:04:09
    combination of the two gives you a
  • 00:04:12
    physically simulated a physically
  • 00:04:15
    grounded Multiverse generator and the
  • 00:04:19
    application the use cases are really
  • 00:04:21
    quite exciting and of course uh for
  • 00:04:24
    robotics uh for industrial applications
  • 00:04:26
    uh it is very very clear this Cosmos
  • 00:04:31
    plus Omniverse plus Cosmos represents
  • 00:04:34
    the Third computer that's necessary for
  • 00:04:36
    building robotic systems every robotics
  • 00:04:39
    company will ultimately have to build
  • 00:04:41
    three computers a robotics the robotics
  • 00:04:44
    system could be a factory the robotics
  • 00:04:45
    system could be a car it could be a
  • 00:04:47
    robot you need three fundamental
  • 00:04:49
    computers one computer of course to
  • 00:04:51
    train the AI we call it the dgx computer
  • 00:04:54
    to train the AI another of course when
  • 00:04:58
    you're done to deploy the AI
  • 00:05:00
    we call that agx that's inside the car
  • 00:05:02
    in the robot or in an AMR or you know at
  • 00:05:05
    the uh in a in a stadium or whatever it
  • 00:05:07
    is these computers are at the edge and
  • 00:05:11
    they're autonomous but to connect the
  • 00:05:13
    two you need a digital twin and this is
  • 00:05:16
    all the simulations that you were seeing
  • 00:05:17
    the digital twin is where the AI that
  • 00:05:20
    has been trained goes to practice to be
  • 00:05:24
    refined to do its synthetic data
  • 00:05:26
    generation reinforcement learning AI
  • 00:05:28
    feedback such and such and so it's the
  • 00:05:31
    digital twin of the AI these three
  • 00:05:33
    computers are going to be working
  • 00:05:34
    interactively nvidia's strategy for uh
  • 00:05:37
    the industrial world and we've been
  • 00:05:39
    talking about this for some time is this
  • 00:05:41
    three computer
  • 00:05:43
    system you know instead of a three three
  • 00:05:46
    body problem we have a three Computer
  • 00:05:48
    Solution and so it's the Nvidia robotics
  • 00:05:51
    the AV revolution has
  • 00:05:53
    arrived after so many years with weo
  • 00:05:57
    success and Tesla's success it has very
  • 00:06:00
    very clear autonomous vehicles has
  • 00:06:02
    finally arrived well our offering to
  • 00:06:05
    this industry is the three computers the
  • 00:06:07
    training systems to train the AIS the
  • 00:06:10
    simulation systems and and the and the
  • 00:06:12
    synthetic data generation systems
  • 00:06:14
    Omniverse and now Cosmos and also the
  • 00:06:16
    computer that's inside the car each car
  • 00:06:19
    company might might work with us in a
  • 00:06:21
    different way use one or two or three of
  • 00:06:23
    the computers we're working with just
  • 00:06:25
    about every major car company around the
  • 00:06:27
    world weo and zuk and Tesla of course in
  • 00:06:30
    their data center byd the largest uh EV
  • 00:06:33
    company in the world jlr has got a
  • 00:06:35
    really cool car coming Mercedes because
  • 00:06:37
    a fleet of cars coming with Nvidia
  • 00:06:39
    starting with this starting this year
  • 00:06:41
    going to production and I'm super super
  • 00:06:43
    pleased to announce that today Toyota
  • 00:06:47
    and Nvidia are going to partner together
  • 00:06:48
    to create their next Generation AVS just
  • 00:06:51
    so many so many cool companies lucid and
  • 00:06:54
    rivan and xiaomi and of course Volvo
  • 00:06:58
    just so many different companies wabby
  • 00:07:00
    is uh building uh self-driving trucks
  • 00:07:02
    Aurora uh we announced this week also
  • 00:07:05
    that Aurora is going to use Nvidia to
  • 00:07:06
    build self-driving trucks autonomous a
  • 00:07:10
    100 million cars build each year a
  • 00:07:12
    billion cars vehicles on a road all over
  • 00:07:15
    the world a trillion miles that are
  • 00:07:17
    driven around the world each year that's
  • 00:07:20
    all going to be either highly autonomous
  • 00:07:23
    or you know fully autonomous coming up
  • 00:07:25
    and so this is going to be a very L very
  • 00:07:27
    large industry I predict that this will
  • 00:07:29
    likely be the first multi-trillion
  • 00:07:31
    dollar robotics industry this IND this
  • 00:07:35
    business for us um notice in just just a
  • 00:07:38
    few uh of these cars that are starting
  • 00:07:41
    to ramp into the world uh our business
  • 00:07:43
    is already $4 billion and this year
  • 00:07:45
    probably on a run rate of about $5
  • 00:07:47
    billion so really significant business
  • 00:07:49
    already this is going to be very large
  • 00:07:51
    well today we're announcing that our
  • 00:07:53
    next generation processor for the car
  • 00:07:56
    our next generation computer for the car
  • 00:07:57
    is called Thor I have right here hang on
  • 00:08:00
    a
  • 00:08:02
    second okay this is
  • 00:08:05
    Thor this is
  • 00:08:07
    Thor this is this is a robotics
  • 00:08:12
    computer this is a robotics computer
  • 00:08:14
    takes sensors and just a Madness amount
  • 00:08:18
    of sensor information process it you
  • 00:08:22
    know een cameras high resolution Radars
  • 00:08:27
    Liars they're all coming into this chip
  • 00:08:29
    and this chip has to process all that
  • 00:08:31
    sensor turn them into tokens put them
  • 00:08:34
    into a Transformer and predict the next
  • 00:08:37
    PATH and this AV computer is now in full
  • 00:08:41
    production Thor is 20 times the
  • 00:08:45
    processing capability of our last
  • 00:08:47
    generation Orin which is really the
  • 00:08:49
    standard of autonomous vehicles today
  • 00:08:51
    and so this is just really quite quite
  • 00:08:53
    incredible Thor is in full production
  • 00:08:55
    this robotics processor by the way also
  • 00:08:57
    goes into a full robot and so it could
  • 00:09:00
    be an AMR it could be a a human or robot
  • 00:09:03
    it could be the brain it could be the
  • 00:09:05
    manipulator uh this Ro this processor
  • 00:09:07
    basically is a universal robotics
  • 00:09:11
    computer the second part of our drive
  • 00:09:14
    system that I'm incredibly proud of is
  • 00:09:17
    the dedication to safety Drive OS I'm
  • 00:09:21
    pleased to announce is now the first
  • 00:09:23
    softwar defined programmable AI computer
  • 00:09:28
    that has been certified if IED up to
  • 00:09:30
    asold D which is the highest standard of
  • 00:09:34
    functional safety for automobiles the
  • 00:09:37
    only and the highest and so I'm really
  • 00:09:40
    really proud of this asold ISO
  • 00:09:43
    26262 it is um the work of some 15,000
  • 00:09:47
    engineering years this is just
  • 00:09:49
    extraordinary work and as a result of
  • 00:09:52
    that Cuda is now a functional safe
  • 00:09:55
    computer and so if you're building a
  • 00:09:57
    robot Nvidia CA
  • 00:10:04
    yep okay so so now I wanted to I told
  • 00:10:06
    you I was going to show you what would
  • 00:10:08
    we use Omniverse and Cosmos to do in the
  • 00:10:13
    context of self-driving cars and you
  • 00:10:16
    know today instead of showing you a
  • 00:10:18
    whole bunch of uh uh videos of of cars
  • 00:10:21
    driving on the road I'll show you some
  • 00:10:23
    of that too um but I want to show you
  • 00:10:25
    how we use the car to reconstruct
  • 00:10:28
    digital twins automatically using Ai and
  • 00:10:32
    use that capability to train future AI
  • 00:10:36
    models okay let's play
  • 00:10:40
    it the autonomous vehicle Revolution is
  • 00:10:44
    here building autonomous vehicles like
  • 00:10:47
    all robots requires three computers
  • 00:10:51
    Nvidia dgx to train AI models Omniverse
  • 00:10:54
    to test drive and generate synthetic
  • 00:10:56
    data and drive agx a supercomputer in
  • 00:11:00
    the car building safe autonomous
  • 00:11:03
    vehicles means addressing Edge scenarios
  • 00:11:07
    but real world data is limited so
  • 00:11:09
    synthetic data is essential for
  • 00:11:13
    training the autonomous vehicle data
  • 00:11:15
    Factory powered by Nvidia Omniverse AI
  • 00:11:19
    models and Cosmos generates synthetic
  • 00:11:22
    driving scenarios that enhance training
  • 00:11:24
    data by orders of
  • 00:11:27
    magnitude first omnimap fuses map and
  • 00:11:31
    geospatial data to construct drivable 3D
  • 00:11:38
    environments driving scenario variations
  • 00:11:41
    can be generated from replay Drive logs
  • 00:11:43
    or AI traffic
  • 00:11:46
    generators next a neural reconstruction
  • 00:11:49
    engine uses autonomous vehicle sensor
  • 00:11:51
    logs to create High Fidelity 4D
  • 00:11:55
    simulation
  • 00:11:56
    environments it replays previous drives
  • 00:11:59
    in 3D and generates scenario variations
  • 00:12:02
    to amplify training
  • 00:12:04
    data finally edify 3DS automatically
  • 00:12:08
    searches through existing asset
  • 00:12:11
    libraries or generates new assets to
  • 00:12:14
    create Sim ready
  • 00:12:18
    scenes the Omniverse scenarios are used
  • 00:12:21
    to condition Cosmos to generate massive
  • 00:12:24
    amounts of photorealistic data reducing
  • 00:12:27
    the Sim to real Gap
  • 00:12:30
    and with text prompts generate near
  • 00:12:33
    infinite variations of the driving
  • 00:12:37
    scenario with Cosmos neotron video
  • 00:12:40
    search the massively scaled synthetic
  • 00:12:42
    data set combined with recorded drives
  • 00:12:46
    can be curated to train
  • 00:12:49
    models nvidia's AI data Factory scales
  • 00:12:53
    hundreds of drives into billions of
  • 00:12:56
    effective miles setting the standard for
  • 00:12:59
    safe and advanced autonomous
  • 00:13:01
    [Music]
  • 00:13:05
    driving is that incredible
  • 00:13:09
    we take take thousands of drives and
  • 00:13:14
    turn them into billions of miles we are
  • 00:13:17
    going to have mountains of training data
  • 00:13:20
    for autonomous vehicles of course we
  • 00:13:22
    still need actual cars on the road of
  • 00:13:25
    course we will continuously collect data
  • 00:13:27
    for as long as we shall live however
  • 00:13:30
    synthetic data generation using this
  • 00:13:33
    Multiverse physically based physically
  • 00:13:36
    grounded capability so that we generate
  • 00:13:39
    data for training AIS that are
  • 00:13:41
    physically grounded and accurate and or
  • 00:13:43
    plausible so that we could have an
  • 00:13:45
    enormous amount of data to train with
  • 00:13:47
    the AV industry is here uh this is an
  • 00:13:49
    incredibly exciting time super super
  • 00:13:52
    super uh uh excited about the next
  • 00:13:54
    several years I think you're going to
  • 00:13:55
    see just as computer Graphics was
  • 00:13:58
    revolutionized such incredible pace
  • 00:14:00
    you're going to see the pace of Av
  • 00:14:02
    development increasing tremendously over
  • 00:14:04
    the next several
  • 00:14:05
    years I I think the next part is is
  • 00:14:10
    robotics so um the chat GPT moment for
  • 00:14:14
    General robotics is just around the
  • 00:14:16
    corner and in fact all of the enabling
  • 00:14:19
    technologies that I've been talking
  • 00:14:20
    about is going to make it possible for
  • 00:14:24
    us in the next several years to see very
  • 00:14:27
    rapid breakthroughs surprising
  • 00:14:28
    breakthroughs in in general robotics now
  • 00:14:30
    the reason why General robotics is so
  • 00:14:32
    important is whereas robots with tracks
  • 00:14:35
    and wheels require special environments
  • 00:14:38
    to accommodate them there are three
  • 00:14:41
    robots three robots in the world that we
  • 00:14:44
    can make that require no green
  • 00:14:47
    fields Brown field adaptation is perfect
  • 00:14:51
    if we if we could possibly build these
  • 00:14:53
    amazing robots we could deploy them in
  • 00:14:56
    exactly the world that we've built for
  • 00:14:58
    ourselves these three robots are one
  • 00:15:01
    agentic robots agentic AI because you
  • 00:15:05
    know they're information workers so long
  • 00:15:07
    as they could accommodate uh the
  • 00:15:09
    computers that we have in our offices is
  • 00:15:10
    going to be great number two
  • 00:15:13
    self-driving cars and the reason for
  • 00:15:15
    that is we spent 100 plus years building
  • 00:15:17
    roads and cities and then number three
  • 00:15:20
    human or robots if we have the
  • 00:15:22
    technology to solve these three this
  • 00:15:25
    will be the largest technology industry
  • 00:15:27
    the world's ever seen
  • 00:15:29
    and so we think that robotics era is
  • 00:15:33
    just around the corner the critical
  • 00:15:36
    capability is how to train these robots
  • 00:15:39
    in the case of human or
  • 00:15:41
    robots the imitation information is
  • 00:15:44
    rather hard to collect and the reason
  • 00:15:47
    for that is uh in the case of car you
  • 00:15:49
    just drive it we're driving cars all the
  • 00:15:50
    time in the case of these human robots
  • 00:15:53
    the imitation information the the human
  • 00:15:56
    demonstration is rather laborious to do
  • 00:15:58
    and so we need to come up with a clever
  • 00:16:00
    way to take hundreds of demonstrations
  • 00:16:03
    thousands of human demonstrations and
  • 00:16:06
    somehow use artificial intelligence and
  • 00:16:10
    Omniverse to synthetically
  • 00:16:13
    generate
  • 00:16:15
    millions
  • 00:16:17
    of
  • 00:16:19
    synthetically generated motions and from
  • 00:16:22
    those motions the AI can learn uh how to
  • 00:16:25
    perform a task let me show you how
  • 00:16:27
    that's done
  • 00:16:35
    General robotics is arriving powered by
  • 00:16:38
    Nvidia Isaac
  • 00:16:43
    Groot okay well let me let me let me
  • 00:16:47
    tell you what I told you I told you that
  • 00:16:49
    we are in production with three new
  • 00:16:54
    Blackwells not only is the grace
  • 00:16:56
    Blackwell supercomputers MV link7 2's in
  • 00:16:59
    production all over the world we now
  • 00:17:01
    have three new Blackwell systems in
  • 00:17:04
    production one amazing AI foundational M
  • 00:17:09
    World Foundation model the world's first
  • 00:17:11
    physical AI Foundation model it's open
  • 00:17:14
    available to activate the world's
  • 00:17:16
    industries of Robotics and such and
  • 00:17:20
    three and three robotics three robots
  • 00:17:23
    we're working on uh agentic AI uh human
  • 00:17:27
    or robots and self-driving cars
  • 00:17:30
    uh it's been an incredible year I want
  • 00:17:32
    to thank all of you for your partnership
标签
  • 物理人工智能
  • Nvidia Cosmos
  • Omniverse
  • 机器人技术
  • 自动驾驶
  • AI训练
  • Thor处理器
  • 数字孪生
  • 合成数据
  • 功能安全