[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-tiangong-humanoid-100m-864-speed-vs-control":3,"topics-all":35,"news-related-10ebf855-6ae7-477c-846f-9783b32fc123":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"10ebf855-6ae7-477c-846f-9783b32fc123","天工跑进 8.64 秒:人形机器人的速度赢了,但步态和制动仍输人类","北京人形机器人创新中心的天工在世界人形机器人运动会上以 100 米 8.64 秒超越博尔特 9.58 秒人类纪录,但 50 多步的步态、起跑延迟与撞软垫的制动暴露电机方案在决策与感知上的不足。","为期五天的世界人形机器人运动会于 8 月在北京揭幕,51 个项目里有 30 项是体育竞技、21 项是场景化竞赛,超过 40% 的项目要求机器人完全自主完成。首日 8 月 22 日的短跑赛事直接把机器比人快这件事摆到了赛场上:两台机器人跑出比博尔特 100 米 9.58 秒人类世界纪录更短的成绩,另一台在 400 米短跑中以 39.7 秒完赛,超越了南非名将 Wayde van Niekerk 43.03 秒的原纪录。相比去年百米仍需 20 秒以上的同款机器人,这是量级层面的跃迁。\n\n## 速度 vs 步态:8.64 秒是怎么跑出来的\n\n北京人形机器人创新中心研制的通用人形机器人天工,被 Reuters 报道在运动会上跑出了 100 米 8.64 秒的成绩,平均时速接近 42 公里\u002F小时,已超过博尔特保持的人类纪录。问题在于它跑成这样的方式非常机器人化——博尔特的起跑反应时间是 0.146 秒,而天工在发令后接近 1 秒才真正迈出第一步;100 米里博尔特只用了 41 步,天工却需要 50 多步,步频明显更高、步幅明显更短,这是电机扭矩曲线和控制策略堆出来的形态,不是肌肉-肌腱-神经系统的产物。韩国光云大学机器人学教授 Park Suhan 直言:这种步态反映的是电机性能,而不是对人类短跑动作的模仿。\n\n## 制动:真正拉开差距的工程难题\n\n更深一层的差距在停下来。现场冲过终点后,机器人没有像短跑运动员那样进入减速曲线,而是以全速直撞十几米开外的巨大软垫,多台翻倒并出现明显损坏——制动不是一个孤立的工程问题,而是感知、决策、机械响应三者协同能力的综合体现。另一名研究者补充了一个反例:如果让机器人迈出更大的步幅,稳定性会迅速崩溃,所以现在的方案本质上是用更密的步子换稳定性。\n\n## 从赛场到工厂:真正要补的课\n\n这场运动会真正在意的并不只是速度。组织方同时设计了模拟工厂、餐厅、办公室、应急场景的任务,要求机器人在处理包装、仓储、线缆接插等任务时,面对角度不当的线缆、刚好够不到的物体、发生位移的包裹等常见工业难题,仍能可靠工作。换言之,跑得快只是一个开端,真正的考卷是能不能像人一样在不那么炫目的岗位上持续工作。专家的判断比较一致:未来几年能转弯、能自主决策的机器人,会比继续压榨短跑成绩更值得追踪。\n\n所以这件事值得读者关注的角度是:8.64 秒和 0.146 秒起跑的差距,本质上是控制循环时延 × 电机响应带宽 × 算法对极限工况的鲁棒性这三个变量的差距。电机能跑出高于人体的峰值速度,但要把这些参数同时顶到在工厂里不出错的水平,还需要感知与决策链路在毫秒级别对齐——这才是从赛场成绩到商用落地的真正鸿沟。\n\n(参考来源:Solidot 引述的赛事报道 https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85169 与 https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85228 ;Reuters 转引原始报道 https:\u002F\u002Fwww.reuters.com\u002Fworld\u002Fasia-pacific\u002Fchinas-record-robotic-strides-show-limits-human-speed-2026-08-28\u002F)","https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85228","d59894d3-308e-4fd8-8865-86dc1eeac4a2",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"86978019-7630-45a7-b417-022f5a124ddb","en","Tiangong clocks 8.64s at 100m: humanoid robots win on pace but lose on gait and braking","Beijing Humanoid Robot Innovation Center's Tiangong ran 100m in 8.64s at the World Humanoid Robot Games, beating Usain Bolt's 9.58s human record. But 50+ strides, a near-1s start delay, and crash-into-cushion braking expose gaps in motor-based control, perception, and decision-making.","The five-day World Humanoid Robot Games opened in Beijing in August with 51 events spanning 30 athletic and 21 scenario-based competitions; over 40% require robots to run fully autonomously. On the first day's track events, robots did what most assumed was still years away: two crossed 100m faster than Usain Bolt's 9.58s human world record, while another finished 400m in 39.7s, beating Wayde van Niekerk's 43.03s mark. Compared to last year's humanoid sprints still stuck above 20 seconds, this is a step-change in capability.\n\n## Speed vs. gait: how 8.64s was actually run\n\nTiangong, a general-purpose humanoid developed by the Beijing Humanoid Robot Innovation Center, ran the 100m in 8.64s at the Games, averaging nearly 42 km\u002Fh — already faster than Bolt's human record according to a Reuters dispatch. The catch is that the way it ran is very much robotic. Bolt's start reaction time is 0.146s; Tiangong needed close to a full second to actually begin moving. Bolt covered 100m in 41 strides; Tiangong needed more than 50 — much higher cadence, much shorter stride. This is the shape produced by torque limits and stability constraints, not a tendon-and-neuron system learning to sprint. Park Suhan, a robotics professor at Korea's Kwangwoon University, put it bluntly: the gait reflects motor performance, not an attempt to imitate human sprinting.\n\n## Braking: the real engineering gap\n\nThe deeper gap is stopping. After crossing the finish line, robots did not decelerate like human sprinters; they charged at full speed into large foam pads placed more than ten meters beyond the line. Multiple robots flipped and visibly broke — braking is not an isolated control problem but a combined perception, decision, and mechanical-response problem. Another researcher pointed out the tradeoff directly: let the robot take a larger stride and stability collapses, which is why today's controllers essentially trade stride length for cadence.\n\n## From track to factory floor: what actually needs fixing\n\nThe Games were never really only about speed. Organizers also staged factory, restaurant, office, and emergency scenarios where robots had to handle packaging, warehousing, and cable-insertion tasks while coping with misaligned cables, barely reachable objects, and shifted parcels — the dull, unpredictable failures humans absorb without thinking. The implicit exam is whether robots can keep working reliably in unglamorous jobs. Experts largely agree: in the next few years, robots that can turn and decide for themselves will be more worth watching than another tenth-of-a-second shaved off the sprint.\n\nThe takeaway angle: the gap between 8.64s and Bolt's 0.146s reaction time reduces to three variables — control-loop latency, motor response bandwidth, and the algorithm's robustness at the limits. Motors can deliver higher peak output than the human body, but pushing all three variables into the regime where a robot doesn't fail in a factory still requires perception and decision loops to align at the millisecond level. That is the real moat between a podium finish and commercial deployment.\n\n(Sources: Solidot relay of the Games coverage https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85169 and https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85228 ; Reuters original dispatch https:\u002F\u002Fwww.reuters.com\u002Fworld\u002Fasia-pacific\u002Fchinas-record-robotic-strides-show-limits-human-speed-2026-08-28\u002F)","tiangong-humanoid-100m-864-speed-vs-control","2026-09-03T00:00:00Z","2026-09-03T07:10:33.901963Z","2026-09-03T07:10:33.901973Z",true,"agent",78,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"82af5716-322e-46cf-9de5-b85e8cdd5712","微软首推安全专用模型 MAI-Cyber-1-Flash:小模型+多智能体,把漏洞挖掘成本砍半","microsoft-mai-cyber-1-flash-mdash","2026-07-28T01:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"a2e8ac5b-ca51-4ddb-88d4-54373d1f0774","SUNTA 用\"惊奇度\"切分视频预测:东京大学让模型在 250 步后仍不崩溃","sunta-surprise-chunking-video","2026-07-04T16:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"e73fe0e6-1b5a-46a2-bbfb-d9a2f1f065d7","AI 智能体挖遍代码库:隐晦式安全在 Patch Tuesday 974 个 CVE 面前失守","ai-agents-kill-security-obscurity","2026-09-15T01:06:05+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"b398dc77-58a1-498e-a9ed-c045c83c90be","AI 抢走消费级 DRAM:一年涨价五倍,手机路由器全被拖下水","ai-dram-consumer-electronics-price-surge","2026-09-14T01:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"e61b1180-f540-4ec5-9796-b24d9258d2ca","AI 辅助挖洞时代的补丁爆炸:微软单月修复 974 个 bug,专家却说「针没变多」","microsoft-974-bugs-ai-haystack","2026-09-11T04:00:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"d17a841b-abca-46e0-80e4-d955f1c837ba","亚马逊 VGT3 仓库曝光:一天拆掉上千本书,只为给 AI 模型喂语料","amazon-vgt3-warehouse-ai-training-books","2026-09-07T03:30:00+00:00"]