新智元报道
老黄又给具身智能打了一针鸡血。
GTC柏林大会上，英伟达官宣扩展Isaac机器人开发平台，模型、数据、仿真工具和计算系统一起上阵。
老黄甚至直接放话，机器人的「ChatGPT时刻」越来越近了。
可问题来了，机器人到底哪天才能真正上岗？
人们盯着灵巧手和机械腿，研发团队还得盯着另一块屏幕：训练任务，怎么又停下来了？
GPU已经备好，数据还在路上。模型等着更新，原始视频还没处理完。等一轮训练好不容易跑起来，多机同步结果的时候，直接又堵死了。
AI很忙，算力却未必一直在有效干活。
就在10月29日至30日，2026腾讯全球数字生态大会在深圳开幕。
在算力与云基座专场，做世界模型、具身大模型和家庭机器人的公司，摆出了各自的探索，也摊开了背后的基础设施需求。
腾讯云副总裁李力在现场说，「没有Infra的AI，就是浮沙之上筑高墙」。
在他看来，AI的智能和决策中枢，终究要回到云上。过去几十年互联网的发展，走的就是这条路：从个人电脑上的单点，走向分层的分布式体系。AI也一样——没人能把最强大的AI一直塞在自己的电脑里，它最终要回归到云上成熟的计算、存储、网络体系当中。
Agent正在加速上云。腾讯云在现场发布全新品牌Agent Hosting，将面向不同场景的云端Agent产品汇聚为一个产品家族，服务个人、开发者和企业，让团队使用有序、投入有据，把复杂的准备做在背后。
李力在致辞中谈到，腾讯云已经做了不少Agent云上托管的尝试。这些探索着眼于如何让Agent在云上稳定运行，让Agent体系及其管理像互联网一样实现工业化。
而具身智能，对云上AI Infra的依赖更深。计算、存储、网络这三项Infra的经典话题，在具身智能这里有了新的命题：
- 计算，决定算力能不能稳定、高效地跑起来；
- 存储，决定海量数据能不能及时喂进训练；
- 网络，决定大规模集群能不能协同好。
本次生态大会上，三家具身智能公司的落地实践，计算、存储两大模块的产品与方案，以及HCCL和AI Link提供的集群通信与网络传输能力，共同呈现了这份「三科合卷」—— 腾讯云AI Infra在具身智能领域的最佳实践。
给AI一张图片，它就能把画面扩展成一个可交互的空间。
按下WASD，可以控制方向、切换视角；用自然语言，也能实时驱动其中的变化。
这是Manifold AI流形空间，带到现场的世界模型成果之一。
在机器人方向，团队也在探索一种更直观的学习方式：让机器人看一次人类示范，再尝试完成新任务。
流形空间创始人、CEO武伟以系鞋带为例，介绍了通过一次示教进行推理的探索。部署中产生的反馈，还能回流用于模型持续改进。
这样的模型，背后是一轮又一轮的大规模训练。
而流形空间选择的，是在国产算力上训。
武伟介绍，双方从算子和基础设施层面联合优化，用了约两周，将相关训练性能提升近3倍。目前，由512颗国产芯片组成的算力集群实现了稳定训练，下一步将持续扩容。
在他看来，过去这类优化大多只能靠团队自己摸索；和腾讯云一起从算子层、Infra层往上调，才让团队对国产算力有了信心。
流形空间的故事，其实说出了很多具身团队都要过的两道坎。
第一道，算力怎么选。如今，越来越多团队在寻求多元化的算力，也在推进国产算力的适配和采用。可换一种芯片，不是换个插头那么简单：框架要兼容，算子要适配，性能还得反复调优。不少团队迁完才发现，模型「能跑，但跑不快」。
第二道，算力怎么用好。算力是一笔不小的投入，多机训练时却常有资源空转；数据处理、仿真、训练的需求此起彼伏，固定配置的集群不是闲置，就是排队。
而具身智能，恰恰把这两道坎都放大了。
大会现场，腾讯云异构计算产品副总监黄阳给了一组数字：现在的机器人只能完成约10%的结构化任务，剩下90%的非结构化任务还没法替代；单一客户的数据量，已经从几十万小时涨到几千万小时，训练需求增长3倍以上；上万台机器人各自试错，群体经验的复用率只有0.01%。
用他的话说，本体的尺度，装不下智能的尺度。所以，具身大脑得长在云上。
黄阳给腾讯云的定位也很干脆：不碰你的算法，只接管你不想管的调度。为此，腾讯云搭起了三大算力底座。
数据底座，把原始数据变成训练集。
采集上来的原始数据，只有约5%能直接用于训练，其余都要先清洗处理。这一层由HAI DataFlow和数据万象承接。
客户自己的算子，可以打包成镜像接进来，业务逻辑不用重写；没有算子，就用数据万象提供的腾讯云自研算子。环境部署、步骤衔接、CPU和GPU怎么配，都交给云，团队只管提交任务、拿结果。
训仿底座，对付的正是流形空间碰到的那两道坎。
所谓算子，可以理解成模型计算里的基本工序。某道工序没适配好，硬件再强也难充分发挥。再加上框架兼容、集群通信和运行稳定性，每个环节都会拖慢进度。腾讯云在这里分三步走：
- 迁得快：HCC多芯算力配合AI算子优化Agent，把模型从原方案迁到国产芯片，迁移速度提升2倍。一家头部具身模型团队的VLA训练任务，两周内完成到国产芯片的迁移，开箱吞吐提升约30%，叠加HyperAcc加速后提升约52%；
- 训得快：HyperAcc训练加速套件一行代码开启，在国产芯片千卡规模实测中，性能提升至少22%；
- 跑得稳：HyperTrace在训练前预检整个集群，训练中旁路采集和监控，停训排障减少91%。
强化学习仿真也跑在同一套底座上，企业保留自己的仿真环境和奖励设计。
推理底座，打通具身应用的「最后一米」。
机器人和它要接触的对象，往往就在一米之内，任务要在百毫秒级完成。
高性能应用服务HAI推出的算力背包，把算力和模型打包在一起，机器人通过订阅接入，用API或SDK调用，不用自己采购和运维算力。AI可信集群拿到了信通院认证，提供全链路加密，数据全程在客户自己手里。
企业专注教会机器人，云负责把算力用好。
算力的问题解决了，下一个瓶颈紧跟着出现：数据。
RoboScience机器科学自研具身大模型Visics，每周新增数十万小时的仿真与视频数据。视频、图像、3D点云、真机回流，格式五花八门；北京、南京两地还各有标注集群。
做家庭机器人的元点Zeroth，则是另一种烦恼：机器人7×24小时在线，视频、语音和传感器日志源源不断，其中大量是重复的无效动作，团队内部管这叫「dirty data」。
一个是数据太杂，一个是高价值数据太难找，指向的却是同一个痛点。
在算法、算力、数据这「模型三要素」里，前两者正在快速收敛：主流算法路线大同小异，算力也越来越多元、越来越容易获得。
所以，腾讯云存储产品解决方案总监、首席技术专家温涛直言，「各家具身公司真正能拉开差距的，是数据的质量、来源和总量，这些决定了具身智能能力的上限」。
可数据从采集到训练，要经过清洗、格式转换、标注、权限与版本管理、交付，任何一环卡住，海量数据就喂不进GPU，算力只能干等。
腾讯云给出的解法，是让数据飞轮转起来：采集、处理、训练，再到真机验证，哪里做得不好，就带着问题开始下一轮。支撑这个飞轮的，是三层架构。
底层：存得下，也存得起
对象存储COS是统一底座。真机、仿真、互联网视频、BAG、LeRobot、MCAP、HDF5等格式都能直接装；单集群可管理百EB级数据，可用性达99.995%，持久性达12个9。
成本同样要算。腾讯云用磁带、QLC、TLC多元介质加冷热分层，该热的放热层，该冷的放冷层，仅COS磁带存储就已累计超过2000PB。
中层：数据流转的「加工管道」
数据万象Robo-Suit提供质检、拆分、标注、重建、检索、脱敏6类算子、100多项能力；MetaInsight支持以文搜图、以图搜图，召回率95%以上。
再加上跨账号、跨园区复制和全球加速，格式转换、权限管控、数据定向交付，都在这一层完成。
顶层：贴着GPU的加速器
对象存储擅长大规模、低成本，和GPU训练之间却隔着一道性能的沟。
数据加速器GooseFS补上了它：单客户端吞吐40Gbps、时延亚毫秒级，热点数据就近缓存、反复读取零回源。新发布的COS Express，读取吞吐达到标准对象存储的10至20倍；高并发、海量小文件的训练场景，则交给全NVMe的并行文件系统CFS Turbo。
三层连起来，标杆客户的实测数据是：数据处理时间减少35%，GPU利用率提升到90%以上，训练时长缩短30%以上，整体存储成本下降30%以上。
再看开头的两家公司。
机器科学创始人兼CEO田野定了一条硬标准：一份数据，多地共享。腾讯云用COS在南京建起统一数据湖，北京、南京跨地域直读同一份数据，省下了两地各建一套高性能存储的数千万元投入；训练侧用CFS Turbo配合400Gbps级RDMA，减少数据等待。
双方还共创了潮汐算力：白天训练，夜里闲置的GPU跑VLM预标注，标注效率提升5到10倍。
元点Zeroth用COS在上海、北美、欧洲多个园区承接PB级数据，智能视图计算平台支撑远程看护的实时查看与回放，MetaInsight帮工程师快速捞出高价值片段。
用联席CTO李元庆的话说：相比一小时重复无效动作，一分钟高密度交互往往更有价值。
数据既要存得下，也要及时读得到。否则，算力再强，也只能等米下锅。
数据送到位，多台机器之间还得配合好。
具身模型越训越大，集群也越拉越大，训练得靠成百上千台机器一起干。单机再强，只要通信慢一拍、网络抖一下，整个集群都得陪着等。
就像一群人合写报告，每人都写得快，汇总时堵在一个窗口，报告照样交不出去。多机配合不好，算力堆得再多也是白搭。
腾讯云给出的答案，是自研的集合通信库HCCL。它配合星脉网络，优化端口流量均衡，并在多任务并发时感知拥塞、动态切换路径。其中，端口均衡优化带来了40%的AllReduce通信带宽收益。
落到具身场景里，一家具身智能公司的真实案例，使用HCCL后，模型端到端性能提升13%。集群之外，具身数据还需要跨地域流转。
一家具身智能客户要把数据在多个节点之间进行调度，原有自建链路难以满足快速扩容需求。
腾讯云AI Link通过多链路负载均衡，组合调度自有骨干网、合作伙伴网络和公网组合，再通过QoS队列分级，将空闲带宽用于大规模数据传输，拥塞时优先保障关键业务。在这一案例中，可交付带宽增至原来的5倍，单位成本降低50%以上。
从集群内协同到跨地域数据传输，两者共同要减少的，都是算力的等待。
从会生成世界的模型，到靠海量数据喂大的具身大模型，再到走进客厅的家庭机器人，流形空间、机器科学、元点三家做的事各不相同，碰到的难处却很像：算力要稳，数据要快，网络要通。
哪一环掉链子，模型就只能干等。
回头看，计算、存储、网络三条线，答的其实是同一道题：具身智能从实验室走向产业，基础设施的复杂度由谁来扛。谁扛得稳，谁就落地得快。
三科合卷，考的不是哪一门的单科成绩，而是计算、存储、网络能不能拧成一股绳。三家具身公司交出的成绩，就是腾讯云AI Infra在具身智能赛道上的最佳实践。
放到整个具身产业来看，竞争的重心也在悄悄转移。过去比的是谁的机器人走得更稳、抓得更准；接下来比的，是谁的数据飞轮转得更快、模型迭代得更勤、能力复制得更广。
本体可以各家自己造，算法可以各家自己写；但从数据采集、训练仿真到部署上线的这条长链路，正越来越多地交给云来承接。
当底座不再拖后腿，机器人从「造出来」到「用起来」的距离，就会被一点点缩短。
机器人真正上岗的那一天，或许并没有想象中那么远。
点击「阅读原文」，直击2026腾讯全球数字生态大会「算力与云基座专场」完整回放。
编辑：桃子
Read more
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Yellow also gave a shot of chicken blood to smart people.
At the GTC Berlin Conference, British officials announced the expansion of the Isaac robot development platform, which brings together models, data, simulation tools and computing systems.
The "ChatGPT Time" of robots is getting closer.
But the question is, when will the robot actually get on duty?
People are staring at smart hands and mechanical legs, and the R & D team has to look at another screen: training missions, why stop again?
GPU is ready and the data is still on the way. The model is waiting to be updated, and the original video is not finished. When the training is hard to run and the results are synchronized, it's blocked.
AI's busy, but it doesn't always work.
Just from 29 to 30 October, the 2026Tencent Global Digital Ecology Congress opened in Shenzhen.
The companies that model the world, have LLM and family robots in the computational and cloud-based arenas set out their own explorations and the underlying infrastructure needs.
Vice-President Tencent Lee Li Li said on the spot that "without Infra's AI, it's the building of walls on the sand."
In his view, AI's intelligence and decision-making hub had to return to the clouds. The development of the Internet over the last few decades has followed this path: from a single point on a personal computer to a hierarchical distributional system. The same goes for AI — no one can stick the strongest AI in their computer, and it eventually goes back to the mature computing, storage, network systems of the clouds.
Agent is speeding up the clouds. Tencent clouds are on the scene under the brand-new publish Agent Hosting, which brings together cloud-based Agent products for different scenarios into a product family that serves individuals, developers and businesses, allowing teams to use orderly and well-documented inputs and to put complex preparations behind them.
In his statement, Li Li said that the Tencent cloud had made many attempts to host it. These explorations focus on how to stabilize Agent in the clouds and industrialization of the Agent system and its management like the Internet.
And smart enough to rely more on AI Infra. The three classic Infra topics of computing, storage, and networking have a new theme here with intelligence:
- Calculating, determining whether or not computing can run in a stable and efficient manner;
- Storage to determine whether big data can be fed in time for training;
- Networks, deciding whether large clusters can work together.
At this ecological conference, the landing practices of three furniture-smart companies, the computation, storage of products and programmes of two main modules, and the cluster communication and network transmission capabilities offered by HCCL and AI Link together present the best practice of Tencent cloud AI Infra in the field of smartness.
Give AI a picture that expands into an interactive space.
Press WASD to control direction, switch perspectives; in natural languages, drive changes in real time.
This is the Manifold AI Flowspace, one of the results of a world model brought to the scene.
In robot direction, the team is also exploring a more intuitive learning approach: show the robot a human demonstration and try to do a new job.
The founder of Fluid Space, CEO Wuwei, gave an example of the exploration of reasoning through a teaching. Feedback from deployment can also be used for continuous model improvement.
This model is followed by a series of large-scale trainings.
And drift-shaped space is chosen by training in national production.
In his presentation, Mu Wei said that the two parties had combined to optimize their skills at the arithmetical and infrastructure levels for about two weeks, increasing their training performance almost threefold. At present, the numerical cluster of 512 nationally produced chips has achieved stabilization training and next step will continue to expand.
In his view, most such optimizations in the past had to be traced by the team itself; together with the Tencent cloud, it was from the arithmetical layer to the Infra layer that the team had confidence in the country ' s productivity.
The story of the fluid space actually tells the two stories the team had to go through.
First, how do you choose the math? Today, more and more teams are seeking to diversify and promote the adaptation and adoption of national production. A different chip is not as simple as a plug: the frame is compatible, the algorithm is compatible and the performance is refined over time. A number of teams moved to find out that the model "can run, but not fast."
Number two, how do you do that? The calculus is a huge input, and resources are often available for multiple training; data processing, simulation and training requirements are running out, and the configuration of clusters is either idle or queued.
And smart enough to magnify both.
On the spot of the Congress, Tencent Deputy Inspector General of Cloud Illustrative Production gave a set of figures: Only about 10 per cent of the current robotic tasks are structured, and 90 per cent of the unstructured tasks remain unsubstituted; the amount of single-client data has increased from hundreds of thousands of hours to tens of millions of hours, and training needs have increased more than threefold; and tens of thousands of robotics have failed each, with only 0.01 per cent of group experience.
In his words, the scale of the body, it doesn't fit the measure of intelligence. So, the brain has to grow on the clouds.
Huang Yang's positioning of the Tencent cloud is also straightforward: without touching your algorithm, taking over the movement you don't want to control. To that end, the Tencent cloud has set up three main algorithms.
Data base, turning raw data into a training set.
Only about 5 per cent of the raw data collected can be used directly for training and the rest must be cleaned and processed first. This layer is carried by HAI DataFlow and the Data Vientiane.
The client's own calculator, which can be packaged into mirrors, does not have to rewrite business logic; without a calculator, the Tencent cloud provided by the data Vientiane is used to study the calculator itself. Environmental deployments, step interfaces, CPUs and GPUs are presented to the clouds, and teams only submit tasks and produce results.
The training of the underseas is dealing with the two canoes encountered in the fluid space.
The so-called algorithms can be understood as basic processes in model calculations. A certain process is inadequate and hardware is not fully available. Together with framework compatibility, cluster communication and operational stability, each link slows down. Tencent clouds walk in three steps:
- Fast-moving: HCC multi-chip calculator optimizes Agent, moving the model from the original program to the national chip, increasing the migration rate twice. The VLA training mission for a head-based model team completed the migration of the nationally produced chips within two weeks, with an increase of about 30 per cent in the opening of the box and an increase of about 52 per cent after the acceleration of HyperAcc;
- Fast-track training: HyperAcc training acceleration package code open, with performance increased by at least 22 per cent in measurements of the size of the national chip kilocalories;
- Steady running: HyperTrace pre-screens the entire cluster before training, collects and monitors sideways in training, and reduces the stop-out barrier by 91 per cent.
Enhanced learning simulations also run on the same floor, with enterprises retaining their own simulation environments and incentive designs.
The bottom seat of the reasoning, the "last metre" of the body.
The robot and the object to which it is exposed are often within one metre and the mission is completed at 100 milliseconds.
High Performance Application Service (HAI) roll-out calculator packs, combining algorithms with models, robots using subscription access, API or SDK, and not purchasing and operating computations. The AI Credible Cluster was certified by the Institute and provided full chain encryption and the data was in the hands of the client itself.
Business is focused on teaching robots, and clouds are responsible for making good calculations.
The problem of calculus is solved, and the next bottleneck follows: data.
RoboScience Machine Science develops its own body, LLMVisics, adding hundreds of thousands of hours of simulation and video data per week. Videos, images, 3D-point clouds, real-time echoes, in a variety of formats; Beijing and Nanjing also have clusters.
Zeroth, a family robot, is another problem: The robots are 7x24 hours online, and video, voice and sensor logs are constantly running, many of which are repetitive and ineffective actions called "dirty data" within the team.
One is that the data is too multiplicity, the other is that the high-value data are too difficult to find, and the other points are directed at the same pain.
In algorithms, algorithms, and data, the three elements of the model, the first two are shrinking rapidly: mainstream algorithms have much the same and much different routes, and computing power is becoming more and more available.
Thus, the Director of Tencent Cloud Storage Products Solutions and Chief Technical Specialist Wing-Tao put it bluntly: "What really opens the gap between furniture companies is the quality, source and volume of data, which determine the upper limit of intellectual ability."
From collection to training, data can be cleaned, format converted, labeled, permission and version managed, delivered, and any ring is stuck, and big data cannot feed to GPU and can only calculate.
The solution given by the Tencent cloud is to turn the data wheel around: gather, process, train, then authenticate, and start the next round with the problem where it does not work. It's a three-tiered structure that supports this wheel.
Bottom: I can save, I can afford.
Object storage COS is a unified base. Realism, simulation, Internet video, BAG, LeRobot, MCAP, HDF5 etc. can be installed directly; 100 EB-level data can be managed in single clusters with 99.995% availability and 12 9 persistence.
The costs are also calculated. Tencent clouds use tape, QLC, TLC multimedia to cool thermal layer, the cool layer, and COS tape storage alone has accumulated in excess of 2000PB.
Middle: "Processing pipe" for data flow
Data like Robo-Suit provides quality checks, split, labels, reconstruction, retrieval, desensitization type six algorithms, more than 100 capabilities; MetaInsight supports literature searches, maps, and over 95 per cent recall rates.
Together with cross-accounts, cross-campaign replication and global acceleration, format conversion, permission controls and data-directed delivery are all completed at this level.
Top level: Accelerator attached to GPU
Object storage is good at large, low-cost, and there is a gap between GPU training.
The data accelerator, GooseFS, supplements it: single client through 40 Gbps, latency submariseconds, hot spot data close to cache, read back zero. COS Express of the new publish, which reads throughput up to 10 to 20 times the standard object ' s storage; training scenarios for high-smoke, small volume files are handed over to CFS Turbo, the parallel file system for NVMe.
Three layers are linked, with measured data for the client of the poles: a 35 per cent reduction in data processing time, an increase in GPS utilization to more than 90 per cent, a 30 per cent reduction in training time and a 30 per cent reduction in overall storage costs.
Look at the first two companies.
The founder of machine science and CEO field set a hard standard: one data, multiple sharing. The Tencent cloud built a unified data lake in Nanjing using COS, and Beijing and Nanjing read the same data directly across the geographical area, saving them tens of millions of yuan in each of the two locations for high-performance storage; training side by side with 400 Gbps grade RDMA to reduce the number of data waiting.
The two sides also created tidal energy: daytime training, with the idle GPU running VLM prescripts, which increased efficiency 5 to 10 times.
The meta-point Zeroth uses COS to host PB-class data in Shanghai, North America and several European parks, smart view computing platforms to support real-time viewing and playback of remote care, and MetaInsight helps engineers quickly extract high-value footage.
In the words of Co-Chair CTO Lee Yuanqing, one minute of high-density interaction is often more valuable than one hour of repetition of invalid actions.
Data must be stored and read in a timely manner. Otherwise, we'll have to wait for rice.
Data to be delivered, multiple machines to fit.
The larger the body model, the larger the cluster, and the hundreds of machines are trained to work together. The single machine is stronger, and the whole cluster will have to wait as long as communications are slow and the network shakes.
Like a group of people co-authoring the report, everyone writes quickly, and the summary is blocked in a window and the report cannot be delivered. It's not working well, it's not working well.
The answer given by the Tencent cloud is HCCL, a collection of self-researched communications. It aligns with the arterial network, optimizes port traffic balance, and senses congestion and dynamical switching paths during multitasking and onset. Of these, the port balance optimization resulted in 40 per cent of AllReduce communications bandwidth gains.
A real case of a furniture-body-smart company, using HCCL, led to a 13 per cent increase in end-to-end performance. Beyond the clusters, there is also a need to move data across geographical areas.
A furniture-based, smart client would have to move data across multiple nodes, and it would have been difficult to build a self-established link to meet rapid expansion needs.
Tencent cloud AI Link uses a multi-chain load balance, a combination of its own backbone, partner network and public network combinations, and a QoS QoS QoS QoS QoS QoS QoS QoS Qo ' s QoS Qo ' s QoS QoS QoS QoS, which is used for large-scale data transmission, giving priority to critical operations when congested. In this case, the deliverable bandwidth increased fivefold and the unit cost decreased by more than 50 per cent.
From intra-cluster synergies to cross-geographical data transmission, both have to be reduced together, and are cost-effective.
From models that generate the world, to large-sized LLMs fed by big data, to family robots that walk into the living room, the fluid space, machine science, meta-points do things differently, but the difficulties are similar: the math is steady, the data is fast, the network works.
If you lose the chain, the model will wait.
In retrospect, the three lines of computing, storage, and network answer is the same question: who is responsible for the complexity of infrastructure, with intelligence going from laboratory to industry. Whoever takes it steady, falls fast.
It's not about which singles, but whether or not to calculate, store, or network into a string. The results of the three furniture companies are the best practices of Tencent cloud AI Infra on a smart track.
The focus of competition is also shifting quietly throughout the physical industry. It used to be more stable and more accurate than whose robots were; then it was who was flying the data more quickly, the models more intuitively, and the ability to replicate more.
It can be built on its own, and algorithms can be written on its own; however, this long link from data collection, simulation training to deployment is increasingly being handed over to clouds.
When the bottom seat is no longer stalling, the distance between "making" and "use" is reduced a little.
The day the robot actually got on duty, maybe not as far as I thought.
The Global Digital Ecology Congress, "Calculation and Cloud-based Grounds" was replayed by clicking on "Reading the original language".
Edit: Peaches
Read more
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.