全部 AI 动态

AK@_akhaliq · 2小时前30

Learning How the World Evolves Extrapolative Video World Models via Latent Dynamics Reasoning paper: https://huggingface.co/papers/2608.09926

译学习世界如何演化 通过潜态动力学推理实现的外推式视频世界模型 论文：https://huggingface.co/papers/2608.09926

Rohan Paul@rohanpaul_ai · 4小时前33

the drunken robot. at the world humanoid robot games in Beijing

译醉酒的机器人。 在北京世界人形机器人运动会上

X.PIN@thexpin · 5小时前40

http://x.com/i/article/2091830460234002432 Weekly Dose of China Tech [08.24.2026] Tencent vs Alibaba's Diverging AI Bets, Xiaomi's Profits Squeezed by Memory Prices, Alibaba's Quiet Chip Weapon, GLM-5.3 Joins Claude and GPT's Tier + One More Thing Hi friends, Hope you had a good week. China’s AI race is entering a new phase. For the past year, the competition was mostly about models: who could build the strongest LLM, who could catch up with OpenAI and Anthropic, and who could achieve frontier performance with fewer resources. But the battlefield is moving. The next question is no longer just who has the smartest AI. It is who can build the infrastructure, finance the spending, and turn AI capability into a sustainable business. Alibaba and Tencent are making two very different bets on that future. One is trying to turn AI into a new infrastructure empire. The other is trying to make AI the next layer of an existing one. At the same time, robotics is facing a similar reality check. China has become one of the world’s most aggressive players in embodied AI, but even the industry’s biggest believers admit that the “ChatGPT moment” for robots has not arrived yet. Technology is moving fast. The harder question is whether the business model can keep up. Anyway, let’s take a look. This Week Features... Tencent vs Alibaba: Two Ways to Bet on China’s AI Future Everyone is watching China’s AI models. Investors are watching something else: the bill. Tencent and Alibaba are both spending billions building their AI futures, but they are making almost opposite bets. Tencent is trying to bring AI into the empire it already owns. Its strategy is built around Hunyuan, WeChat agents, and productivity tools like WorkBuddy. The company’s biggest advantage is not the model itself, but the ecosystem around it: hundreds of millions of users, businesses, payments, advertising, and one of China’s most valuable consumer platforms. Alibaba is taking the opposite path. Instead of adding AI to an existing business, it is rebuilding itself around AI infrastructure: chips, cloud computing, foundation models, and enterprise services. The difference matters because the AI race may not be won by whoever builds the best chatbot. It may be won by whoever owns the infrastructure layer underneath every chatbot. But both strategies come with a difficult question. Tencent needs to prove AI can create a new growth engine inside an existing empire. Alibaba needs to prove that massive AI spending today can become tomorrow’s advantage rather than tomorrow’s cost. This week’s feature looks at the two biggest bets in China’s AI race — and the very different assumptions investors are being asked to finance. [Read the full piece →] The News… (I) Unitree’s CEO Says Humanoid Robots Aren’t Ready Yet Despite the Hype The humanoid robot market has one uncomfortable problem: investors are moving faster than the technology. At the World Robot Conference, Unitree founder Wang Xingxing tried to lower expectations, saying humanoid robots are still not ready for widespread factory or home use. The reason is not hardware. It is adaptability. A robot that can perform a fixed demonstration is very different from one that can walk into an unfamiliar environment and complete a new task without extensive retraining. Wang’s benchmark for a true “ChatGPT moment” in robotics is ambitious: robots should complete roughly 80% of tasks in 80% of unfamiliar environments through simple voice or text commands. He expects that breakthrough within 3–5 years. The market, however, is already pricing in that future. Unitree earned around $41 million in profit on $252 million revenue in 2025, but its IPO valuation implies roughly 219x earnings. The robot revolution may happen. The question is whether it arrives before investors run out of patience. (II) Xiaomi’s Profits Take a Hit From Memory Chip Prices Xiaomi’s H1 2026 numbers show the squeeze: revenue down 8.4% YoY to $28.9 billion, adjusted net profit down 42.8% to $1.7 billion. Still, Q2 alone beat Wall Street’s expectations, pulling in $15.1 billion in revenue and $790 million in profit. The culprit is memory price inflation, which dragged Q2 phone revenue down 7.5% even as Xiaomi held gross margin at a respectable 8.5%. Management is betting things improve once memory prices normalize, but costs are staying elevated for now. AI spend isn’t slowing down: Xiaomi poured $2.5 billion into R&D, with AI eating up nearly 30% of that. Its in-house model, Mimo, is still playing a supporting role powering HyperOS, smart devices, and eventually autonomous driving rather than being monetized directly. EVs remain the growth story, though Xiaomi looks likely to fall short of its 550,000-unit annual delivery target, currently tracking around 35,000 units a month. Overseas EV expansion is now penciled in for H2 2027, starting with Europe. Read Full Article

译腾讯与阿里在AI战略上押注相反路径：腾讯依托混元、微信智能体与WorkBuddy等工具将AI融入现有生态，阿里则围绕芯片、云计算和基础模型重建AI基础设施。宇树CEO王兴兴称人形机器人尚不成熟，预计3-5年内实现"ChatGPT时刻"。小米H1 2026营收同比降8.4%至289亿美元，受内存芯片价格挤压利润。GLM-5.3已加入Claude和GPT所在的第一梯队。

Generalist@GeneralistAI · 6小时前29

We've reduced the time it takes to go from physical prompt → robot behavior. The faster anyone can teach a robot to do something new, the easier it becomes to scale physical work. Read more about GEN-1.5 in our blog post in the comments below.

译我们缩短了从物理提示到机器人行为所需的时间。 任何人教会机器人新技能的速度越快，规模化物理工作就越容易。 更多关于 GEN-1.5 的内容，请阅读下方评论中的博客文章。

X.PIN@thexpin · 11小时前51

Unitree Robotics shares fell more than 10% intraday, erasing about $28 billion in market value from its IPO-day peak. The company opened with a market capitalization of RMB 444.9 billion ($62 billion) after its STAR Market debut. Instead of launching a new robot, Unitree introduced a 7-axis bionic robotic arm priced from RMB 9,900 ($1,380), targeting sorting, assembly and research applications. Founder Wang Xingxing continues to lower expectations, saying the “ChatGPT moment” for embodied AI may still be five years away. The challenge is no longer building robots, but making them truly useful.

译宇树科技（Unitree Robotics）股价盘中跌超10%，市值较IPO首日峰值蒸发约280亿美元。该公司科创板上市首日开盘市值达4449亿元人民币（620亿美元）。宇树未发布新机器人，而是推出起售价9900元（1380美元）的七轴仿生机械臂，面向分拣、装配及科研场景。创始人王兴兴持续降低预期，称具身智能的"ChatGPT时刻"可能还需五年。

X.PIN@thexpin · 12小时前54

China’s first Humanoid Robot Games delivered a surprising result: robots are fast — but not always in control. At the Beijing event, Lightning Robot from Honor finished the 100-meter race in 9.47 seconds, while Tiangong Robot clocked 9.39 seconds, both beating Usain Bolt’s 9.58-second world record. But there was a catch: the robots struggled to stop. They crashed into safety barriers, and Tiangong Super Power lost control and ran toward the audience. Leihe collapsed and had to be carried away.

译中国首届人形机器人运动会结果出人意料：机器人速度飞快，但控制力不足。北京赛场上，荣耀的闪电机器人百米跑出9.47秒，天工机器人以9.39秒完赛，双双超越博尔特9.58秒的世界纪录。但机器人难以刹停，撞上安全护栏，天工超能甚至失控冲向观众，雷鹤倒地后需被抬离。

赵纯想@chunxiangai · 19小时前37

我认知太低了。以下复述大哥原话： 1、宇树市值有 85% 的国家意志，具身智能行业龙头，如果市值比一家互联网公司还低，那机器人这个领域中国就不要搞了。光伏、新能源汽车、机器人，这三个是台柱子。演员不化妆，戏就别演了。 2、还有 15% 是赖子优势。咱们就说他是造玩具，没错，谁都知道他是造玩具。就造。把玩具供应链打磨好了，然后像个赖子一样赖着。我也不研发世界模型，我一共就 5 个博士生，也不搞 lab，就百万年薪养着。世界模型那边只要哪个方向一突破，赖子王公司第二天量产，第三天换脑子，第四天就上工厂了。供应链比你快十倍。

译赵纯想转述观点称，宇树市值含85%国家意志，作为具身智能龙头，其市值不应低于互联网公司，否则机器人产业难以为继。另15%为供应链优势：不研发世界模型，靠5个博士生百万年薪养着，待世界模型方向突破后，以快十倍的供应链速度跟进量产。

Rohan Paul@rohanpaul_ai · 23小时前38

They own the track and the high bar, but the penalty box still belongs to the humans, at least for sometime. posture control after contact is such a hard problem. Autonomous 5v5 soccer is underway at the World Humanoid Robot Games, with the machines recovering from falls more often than earlier trials but still far from controlled possession. https://x.com/KanekoaTheGreat/status/2091554941454397742/video/1

译他们掌控着赛道和高标准，但禁区仍属于人类，至少暂时如此。 接触后的姿态控制是一个极其困难的问题。 世界人形机器人运动会上的自主5v5足球赛正在进行中，机器人们从摔倒中恢复的次数比早期试验更多，但仍远未实现有控的控球。

Rohan Paul@rohanpaul_ai · 1天前26

they look completely at home on that track. at under 20 kg and suitcase-friendly height, these Booster K1s are very portable. smaller size and lower mass make it safer around people

译它们在那条轨道上看起来完全如鱼得水。 重量不到 20 公斤，高度适合登机箱，这些 Booster K1 非常便携。更小的体积和更轻的质量使其在人群周围更安全。

DogeDesigner@cb_doge · 1天前39

BREAKING: Elon Musk's Neuralink is collaborating with Neko Health to identify potential candidates for an upcoming patient round. With explicit consent Neuralink’s dedicated eligibility survey can use Neko scan and medical data to simplify its screening.

译突发：埃隆·马斯克的 Neuralink 正与 Neko Health 合作，为即将开展的一轮患者招募筛选潜在候选人。 在获得明确同意后，Neuralink 专门的资格调查可利用 Neko 的扫描和医疗数据来简化其筛选流程。

Rohan Paul@rohanpaul_ai · 1天前39

we are crossing a threshold. These full-size robots, at World Robot Conference in Beijing, are marketed for emotional companionship, elderly care/assistance. • replicates up to ~90% of fundamental human movements (sitting, leaning, lying down, hugging, basic walking on flat indoor surfaces). • supports >30 complex micro-expressions (blinking, eyebrow raising, smiling, lip pursing, etc.). • Medical-grade / platinum silicone biomimetic skin with gel structure. Replicates pores, skin texture, subcutaneous vessels/veins, fingerprints, and warm/human-range surface temperature. • electronic skin with multi-dimensional sensors across the body for touch and hug detection. • have 88 degrees of freedom (servo joints). • walking speed (Pro/related claims): ~4 km/h typical, up to ~5 km/h max. https://x.com/Le_Figaro/status/2090861861973471320/video/1

译北京世界机器人大会上展出的全尺寸人形机器人，主打情感陪伴与老年人护理，可复现约90%的人类基础动作（坐、躺、拥抱、室内平地行走），并支持超30种复杂微表情。其医用级铂金硅胶仿生皮肤可模拟毛孔、纹理、皮下血管与指纹，全身电子皮肤配备多维触觉传感器，拥有88个自由度，行走速度约4 km/h、最高约5 km/h。

Rohan Paul@rohanpaul_ai · 1天前31

New Stanford and Peking University paper says train a robot policy inside a learned world model and it inherits every mistake the model makes. Those errors pile up as tasks get longer and images get messier. QWM (Q-LEARNING WITH WORLD MODELS), never trains anything inside the model. The world model never touches training here, it only helps the robot choose. – arxiv. org/abs/2608.17163 Title: "Q-Learning With World Models"

译斯坦福大学和北京大学的新论文指出，在学到的世界模型内部训练机器人策略，会继承该模型的每一个错误。 随着任务变长、图像变复杂，这些错误会不断累积。 QWM（基于世界模型的 Q 学习）从不在模型内部训练任何内容。 在此方法中，世界模型完全不参与训练，它只帮助机器人做选择。 - arxiv.org/abs/2608.17163 标题："Q-Learning With World Models"

Rohan Paul@rohanpaul_ai · 1天前19

speed is solved, stopping is not. crash, sparks, fire, repeat.

译速度问题已解决，停止才是难题。 碰撞、火花、起火，循环往复。

Rohan Paul@rohanpaul_ai · 1天前50

This is bigger than a robot race. History just got rewritten. A humanoid robot from Beijing clocked 9.39s over 100m, faster than Usain Bolt’s 9.58s record from 2009. And this year's 100m was autonomous-only, i.e. the robot itself had to complete the race without human steering. https://x.com/Reuters/status/2091244010690212206/video/1

译这不仅仅是机器人竞赛。历史刚刚被改写。来自北京的人形机器人以9.39秒跑完100米，快于博尔特2009年9.58秒的纪录。 而且今年的百米赛是纯自主模式，即机器人必须在不经人类操控的情况下完成比赛。

Rohan Paul@rohanpaul_ai · 2天前23

The running part was pure mechanical poetry. An then that robot just rewrote the definition of quitting while winning.

译奔跑的部分纯粹是机械的诗意。 然后那个机器人直接改写了"在胜利时退出"的定义。

Rohan Paul@rohanpaul_ai · 2天前35

An affordable version of this could change daily life for so many seniors A robot guide dog just led a reporter through a mock busy airport terminal at Beijing’s 2026 World Robot Conference, finding check-in and the right gate.

译一个价格亲民的版本可能会改变许多老年人的日常生活 在北京2026世界机器人大会上，一只机器人导盲犬带领一名记者穿过了模拟的繁忙机场航站楼，成功找到值机柜台和正确的登机口。

Rohan Paul@rohanpaul_ai · 2天前39

hard to look away from the block the group feels unstoppable Booster Robotics brought 80 T2 platforms to the 2026 World Humanoid Robot Games opening in Beijing.

译很难将目光从这块场地上移开 这支队伍让人感觉势不可挡。宇树科技（Booster Robotics）携80台T2平台亮相2026年世界人形机器人运动会北京开幕式。

Rohan Paul@rohanpaul_ai · 2天前30

🇨🇳 Chinese rescue robot: autonomous flight to the drowning person. flies at roughly 10–14 m/s (about 30 mph peak), covers up to ~2 km / 1.2 miles, lands to provide flotation for up to two 80 kg adults, then returns on its own.

译🇨🇳 中国救援机器人：自主飞向溺水者。飞行速度约 10-14 米/秒（峰值约 30 英里/小时），覆盖范围约 2 公里/1.2 英里，降落为最多两名 80 公斤成年人提供浮力，随后自主返回。

Rohan Paul@rohanpaul_ai · 2天前36

the "do-not-touch" rule suddenly makes perfect technical sense 😄 2026 World Robot Conference in Beijing shows an actual UBTECH ultra-bionic humanoid (semi-torso model) with highly realistic silicone skin ~ $17K and for companionship use.

译"不可触碰"规则突然变得在技术上完全合理了😄 2026年北京世界机器人大会展示了一款优必选超仿生人形机器人（半躯干型号），配备高度逼真的硅胶皮肤 约1.7万美元，用于陪伴用途。

Rohan Paul@rohanpaul_ai · 2天前36

pure mechanical carnage. during a pre-race test for the 2026 Beijing Yizhuang humanoid robot half-marathon

译纯粹的机械混战。 发生在2026北京亦庄人形机器人半程马拉松的赛前测试期间。

Rohan Paul@rohanpaul_ai · 2天前43

the guy in the suit looked more nervous UBTECH robot’s live ballroom dance at 2026 World Robot Conference, Beijing. Real-time ankle micro-adjustments maintain even force on both feet against an unpredictable human dance partner.

译穿西装的那个人看起来更紧张 优必选机器人在2026年世界机器人大会（北京）现场表演交谊舞。 实时脚踝微调让机器人在面对不可预测的人类舞伴时，双脚保持均匀受力。

Rohan Paul@rohanpaul_ai · 3天前13

wonder how much energy storage is in those actuators and that constant foot strike frequency The human was trying though, but he never had a chance once it started.

译想知道那些执行器里有多少能量储存 还有那恒定的脚步频率 人类虽然努力了，但一旦它启动，就再无机会。

Rohan Paul@rohanpaul_ai · 3天前40

A lot of embodied AI still feels like AI modules bolted onto a robot. TARS is taking a different architectural bet with AI World Engine (AWE) 3.5, TARS’ embodied-native foundation model for physical AI. Its "Born as One" approach puts action, perception, geometry, and touch into one model from the beginning rather than stitching those capabilities together later. The same model-driven system is designed to generalize across different tasks, objects, environments and robot bodies. The training recipe then implements and validates a full closed-loop methodology for embodied-native foundation models through pre-training and post-training. During pre-training, 2 priors give the model a base understanding of action patterns, spatial structure and understanding of physical laws before it is adapted to a robot, while post-training uses the AI World Engine to roll possible future states forward inside the model, predict what different actions may lead to and use those predictions to choose better actions. TARS describes the full loop as 5 connected parts: embodied-native architecture, dual-prior pre-training, World Engine-driven post-training, scaling validation and continuous data feedback. TARS positions AWE 3.5 as one of the most powerful embodied-native foundation models for general-purpose physical AI, with several minutes of long-horizon closed-loop reasoning and roughly 2x task execution efficiency versus PI0.5. @TARSRobotics #AWE35 #TARS #tarsrobotics 🧵 1.

译TARS 推出具身原生基础模型 AWE 3.5，采用"Born as One"架构，将动作、感知、几何与触觉整合进单一模型，而非后期拼接。该模型支持数分钟长时程闭环推理，任务执行效率约为 PI0.5 的 2 倍，并通过预训练双先验与后训练世界引擎滚动预测来优化决策。

Jim Fan@DrJimFan · 3天前72

The sense of touch is the most criminally under-explored modality in robotics. Imagine doing sleight of hand wearing thick oven mitts. That's exactly how a robot feels today if it were alive. A magnetic piece snapping into place, a paper cup peeling out of a stack, a USB negotiating its way into the port - all invisible to the camera. Learning how to feel must be a full-stack co-designed effort. We are open-sourcing a principled methodology called "T-Rex": 1. Tactile as first-class citizen of the model. Our mixture-of-transformer runs two clocks asynchronously: a slow visuomotor expert plans the motion, and a fast tactile expert refines it in real time with high-frequency corrections at 4 "touch ticks" per vision tick. Forces change faster than frames arrive, so the architecture had to as well. 2. Open data. The largest tactile dataset ever released to our knowledge: a 50-hour (~5,500 episodes) high-quality, carefully synchronized robot play corpus, collected on SOTA tactile hand hardware with 22 degrees of freedom. Available today on HuggingFace! 3. Training recipe: T-Rex extends our prior work, EgoScale. Human egocentric videos for pretraining, a diverse dose of tactile robot play for mid-training. Our experiments show this bridges contact-free pretraining to contact-rich manipulation remarkably well. Pixels are cheap and everywhere, but they run out of steam at the moment of contact. Tactile will carry the last mile. The next scaling curve will be measured in hours of touch. T-Rex is a great collaboration between NVIDIA and Berkeley: 🧵

译NVIDIA 与伯克利联合开源触觉机器人方法 T-Rex，将触觉作为模型一等公民，采用双时钟异步架构，以每视觉 tick 4 次触觉 tick 的高频修正实时优化操作。

Rohan Paul@rohanpaul_ai · 3天前38

10,000 parcels. 5 hours, 14 minutes. One embodied AI model running the whole challenge. X Square Robot's WALL-B model sustained 1,911 parcels/hour, or about 1.88 seconds/parcel, while sorting 10,000 packages. The task is deceptively physical. The arm has to identify package orientation, flip each parcel label-side up, then slide it onto the conveyor. Odd objects such as soft toys are routed separately. For context, Figure has reported a 2.88-second parcel cycle time. X Square Robot's WALL-B model decides how each parcel should be handled from the scene, while its six-axis arms execute the picks and corrections. Its previous public run averaged 1,816 parcels an hour with over 98% accuracy, so the new result claims both longer duration and higher throughput. The apparent simplicity of picking a parcel hides a repeated closed-loop computation spanning 3D perception, grasp reasoning, motion planning, force control, and online correction. The difficulty comes from combining fast visual perception, 3D geometry, grasp selection, collision-aware motion planning, feedback control, and failure recovery under a scene that changes after every action. Sustaining the full loop across 10,000 parcels tests whether the robotics stack remains stable when thousands of small physical uncertainties accumulate. Another point of difficulty comes from the endurance, because this is as difficult as the speed. Every additional hour gives perception errors, tiny calibration errors, grasp failures, awkward package geometries, and small control mistakes more opportunities to compound. A system that looks great for 50 parcels can behave very differently after several thousand. The footage also shows the policy doing more than repetitive pick-and-place. Packages are flipped label-side up, moved onto the conveyor, and unusual items such as soft toys are sent to another lane.

译X Square Robot 的 WALL-B 具身智能模型以 1，911 件/小时的速度完成 10，000 件包裹分拣，全程耗时 5 小时 14 分钟。该任务需实时识别包裹朝向、翻转至标签朝上并滑上传送带，软玩具等异形物单独分流。相比 Figure 报告的 2.88 秒/件周期，WALL-B 在更长时长下实现更高吞吐。

Rohan Paul@rohanpaul_ai · 3天前51

Humanoid robotics has become a capital race, with 2026 VC funding already at $8.7 B and much of it concentrated among a small group of companies. Total VC money is flooding into humanoid robotics: ($8.7B) nearly 2x 2025’s record. Neura Robotics grabbed the biggest slice with a $1.4 billion Series C. The sector’s competitive map is changing on two fronts at once, as venture capital surges and the race’s center of gravity moves toward China. Chart from dealroom

译人形机器人已成为一场资本竞赛，2026年风险投资额已达87亿美元，且大部分集中在少数几家公司手中。 涌入人形机器人的风险投资总额：87亿美元，几乎是2025年纪录的两倍。 Neura Robotics以14亿美元的C轮融资拿下最大份额。 随着风险投资激增，以及竞赛重心向中国转移，该行业的竞争格局正在两条战线同时发生变化。 图表来自dealroom

Rohan Paul@rohanpaul_ai · 3天前25

That’s why fingers matter so much for humanoid robots, and is such a hard problem. The bag was in the hand for maybe one second

译这就是为什么手指对人形机器人如此重要，也是一个如此困难的问题。 袋子在手里只停留了大约一秒钟。

X.PIN@thexpin · 3天前44

China’s robotics market is accelerating. According to the Ministry of Commerce, sales of embodied AI robots on major platforms rose 95.1% YoY in July, while exoskeleton devices grew 39.5%, action cameras 24.7%, and robot vacuum cleaners 19.4%. Robot exports are also expanding rapidly. In the first seven months of 2026, China exported $1.0 billion worth of industrial robots, up 13.2% YoY. In H1, cleaning robots and intelligent bionic robots combined exports reached RMB 18.09 billion $2.5 billion, with industrial robots shipped to 141 countries and regions.

译中国机器人市场加速增长，商务部数据显示7月具身智能机器人线上销量同比增95.1%，外骨骼设备增39.5%。2026年前7个月工业机器人出口额达10亿美元，同比增13.2%；上半年清洁机器人与智能仿生机器人合计出口180.9亿元人民币，工业机器人已出口至141个国家和地区。

DogeDesigner@cb_doge · 3天前28

ELON MUSK: Robot doctors could become extremely competent and deliver care so advanced that everyone on Earth receives better treatment than the best human doctors today.

译埃隆·马斯克：机器人医生可能会变得极其能干，提供的医疗服务将如此先进，以至于地球上的每个人都能获得比当今最优秀的人类医生更好的治疗。

X.PIN@thexpin · 3天前28

China isn’t just hosting the World Robot Conference—it’s also holding the World Humanoid Robot Games. The gray runner is from HONOR. If it starts sprinting, stay out of its way. The white robot looks late for something: it keeps wiping its face mid-run.

译中国不仅主办世界机器人大会，还在举办世界人形机器人运动会。灰色跑者来自荣耀（HONOR）。如果它开始冲刺，请离它远点。白色机器人看起来像是要迟到了：它跑步途中一直在擦脸。

Rohan Paul@rohanpaul_ai · 3天前31

she holds the stage better than expected. At World Robot Conference, Elf-Xuan 2.0 from AheadForm delivered a live vocal performance with real-time lip and expression sync plus arm gestures.

译她在舞台上的表现比预期更出色。 在世界机器人大会上，来自宇树（AheadForm）的 Elf-Xuan 2.0 进行了现场演唱，实现了实时嘴唇与表情同步，并伴有手臂动作。

Rohan Paul@rohanpaul_ai · 3天前19

Another angle of that viral video. residual kinetic energy found the joint’s weak axis. waist linkage was the limit.

译那条爆火视频的另一个角度。 残余动能击中了关节的薄弱轴。 腰部连杆成了极限。

DogeDesigner@cb_doge · 3天前35

“Optimus will actually eliminate poverty. Optimus will actually give people incredible medical care. Optimus will ultimately be better than the best human surgeon with a level of precision that that isn't possible, that is beyond human.” — Elon Musk

译"Optimus 将真正消除贫困。Optimus 将真正为人们提供不可思议的医疗服务。Optimus 最终将超越最优秀的人类外科医生，其精准度是人类无法企及的。" - 埃隆·马斯克

DogeDesigner@cb_doge · 3天前37

ELON MUSK: "If you've got humanoid robots that have very high dexterity and are incredibly smart, it means that everyone on earth will have access to better medical care, than the richest person on Earth. I had to have like a neck surgery three times because the first two ones were done wrong. Back pain may be one of the things that it eliminates. Average happiness level for humans would just go upstream tremendously."

译埃隆·马斯克："如果人形机器人拥有极高的灵活性和惊人的智能，那么地球上的每个人都将获得比当今最富有的人更好的医疗服务。 我不得不做三次颈部手术，因为前两次都做错了。背痛可能是它能消除的问题之一。人类的平均幸福水平将大幅提升。"

Elon Musk@elonmusk · 4天前26

One day, Optimus + Grok will provide incredible medical care to all the people of Earth

译总有一天，Optimus + Grok 将为地球上所有人提供不可思议的医疗服务。

AI at Meta@AIatMeta · 4天前53

Muse Spark 1.2 supports a broad range of multimodal tasks, from turning visuals into working code to translating perception into physical action. It also brings robust audio-visual understanding to enable video-heavy workflows common in real-world enterprise use. Today, we’re sharing new evals and demos that illustrate the breadth of the model’s visual understanding and reasoning capabilities. Let’s start with a demo that shows how Muse Spark parses multimodal observations and calls tools to guide a robot to navigate in an unstructured environment to find a rubber duck. 🧵👇

译Meta 发布 Muse Spark 1.2 新评测与演示，展示其多模态能力广度，涵盖视觉转代码、感知转物理动作及音视频理解。演示中，模型解析多模态观察并调用工具，引导机器人在非结构化环境中导航寻找橡皮鸭。

🚨 AI News | TestingCatalog@testingcatalog · 4天前34

Humyn Labs opened an "egocentric" video library for physical AI! 5 first-person datasets by environment and capture type 6-DoF head poses, 21-point hand keypoints Absolute metric depth + per-frame action labels Delivered in MCAP, RLDS and LeRobot v3 Preview below 👀

译Humyn Labs 推出面向物理 AI 的"自我中心"视频库，包含 5 个以上第一人称数据集，按环境和采集类型分类。数据提供 6-DoF 头部姿态、21 点手部关键点、绝对度量深度及逐帧动作标签，以 MCAP、RLDS 和 LeRobot v3 格式交付。该公司旨在将人类经验转化为机器人技能，弥补机器人数据缺乏互联网级捷径的短板。

Jim Fan@DrJimFan · 4天前47

Seeing a hype wave around GEN-1.5, and rightfully so. Lots of respect to Pete & Andy for executing so well. The secret is in the naturally repetitive motions in human-collected data. There're 2 main sources for such repetitions: (1) Symmetric patterns. Sorting, tidying, and assembling almost never finish in one motion. Open any assembly manual from IKEA, and you find most objects symmetrical. You drive one bolt, then its twin, then the next pair. Every {bolt A, bolt B} pair is a natural continuation in context, and the second instance is a free training signal that imitates the first ("prompt"). (2) Recovery. Humans drop things all the time, but we pick them up so fast, we don’t even notice. That reflex to fix is half of our physical competence. The key insight is to keep the failed first half instead of trimming it away. If the model consumes the full arc, fumble, catch, continue, then recovery shows up organically at test time. It's funny that in-context improvement results from *NOT* over-sanitizing your data. The other critical ingredient is UMI. I've been saying for a while that teleop will not last, and GEN-1.5 is driving the final nail in the coffin. UMI is essentially a human wearing the robot gripper to collect data directly (human → data). Teleop inserts a layer of separation: human → VR/skeletal device → robot → data, which bleeds out all the human "physical intuition". The subtle sleight of hand we perform constantly with objects, the micro-adjustments, the feel of a part snapping into place, is nearly impossible to capture when you can't feel the environment directly. Once you have enough data, many behaviors can actually be zero-shot. For example, you don't even need finetuning to pick up a novel object. The model "just knows" what to do given a similar scene in the training distribution. Whether in-context learning truly works or not also depends on how far away the test is from training. Currently, the demos are still a bit too simple to conclude. I'm cautiously optimistic. Still, it's a great day in robotics.

译Jim Fan 称赞 GEN-1.5 的成功，认为关键在于人类数据中自然的重复动作，包括对称模式与失误后的恢复行为。他强调 UMI 直接采集人类数据优于遥操作，能保留物理直觉。他持谨慎乐观态度，认为当前演示仍过于简单，尚不能断言上下文学习已真正奏效。

X.PIN@thexpin · 4天前45

At the World Robot Conference: Cruella is in Beijing—with her Dalmatian. Meet BoBo, Vbot’s $1,930 family robot dog. It follows by sight, responds to leash tugs and carries snacks. Our tests found a 5-hour battery, solid trail performance, quiet shoes and live video calls.

译在世界机器人大会上：Cruella 来到了北京--带着她的斑点狗。认识一下 BoBo，Vbot 售价 1，930 美元的家庭机器狗。它能通过视觉跟随，对牵引绳的拉动做出响应，还能携带零食。我们的测试发现它有 5 小时续航、扎实的越野表现、安静的鞋底和实时视频通话功能。

Rohan Paul@rohanpaul_ai · 4天前31

LLMs got the internet. Robots have to build their own internet. That is the data problem Humyn Labs is going after. @humynlabs is turning human experience into synchronized training data that robotics cannot readily scrape from the web. A useful robotics dataset cannot just be hours of first-person video. Humyn's samples pair human activity with signals such as IMU (inertial measurement unit), stereo depth, 6-DoF head pose, 21-point hand keypoints, wrist tracking, object tracking and dense action labels. Some captures even synchronize a head camera with both wrist cameras and separate IMU streams. So Humyn is trying to preserve enough structure around those human-demonstrations to make them useful: egocentric video, inertial measurements, hand and head pose, object trajectories, depth, narration and synchronized multi-camera views.

译Humyn Labs 正将人类经验转化为机器人无法从网络获取的同步训练数据，其样本配对 IMU、立体深度、6-DoF 头部姿态、21 点手部关键点、手腕追踪、物体追踪及密集动作标签等多模态信号。与从整个互联网学习的 LLM 不同，机器人数据缺乏捷径，需真实世界经验，该公司正构建 Physical AI 学习的数据平台，涵盖四种感官模态。