[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nvidia-long-wam-robot-context":3,"topics-all":38,"news-related-376751a1-de8e-4e5b-afdc-4e26d7574146":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"376751a1-de8e-4e5b-afdc-4e26d7574146","NVIDIA Long-WAM：机器人记住 19 秒，预测未来再出手","NVIDIA 发布 Long-WAM 世界-动作模型：机器人先读 19.2 秒视觉历史，在潜空间预测未来再出动作。上下文从 2.4 秒扩到 19.2 秒，RoboCasa 成功率 66.3% 升至 78.7%；真机动态叠杯 19\u002F20，两个对比基线 0\u002F20。RTX 5090 每动作块 107.4 毫秒。","实时机器人控制一直有个两难：要判断物体运动和任务进度，策略得看足够长的视觉历史；可历史一长，推理变慢，动作就跟不上物理世界。NVIDIA 研究团队 10 月 7 日提交到 arXiv 的 Long-WAM 给出的答案是模型与系统一起改：把世界-动作模型的视觉上下文扩到 19.2 秒，同时用一整套推理优化把单步延迟压进百毫秒级。\n\n## 记忆越长越好，但红利只属于自回归预训练\n\n核心发现：看得见历史，不等于会用历史。它的视频专家从 LongLive2.0-Robot 检查点继续自回归训练，数据是约 1 万「窗口等价小时」的机器人与第一视角视频（RoVid-X、AgiBot World、EgoDex、EgoVerse、VITRA），不带动作标签，做最长 30 秒的下一块预测；之后经不对称注意力接口（Mixture of Transformers）与动作专家耦合，视频查询永远不读动作 token。\n\n收益写在数字里：RoboCasa GR-1 上，观察上下文从 2.4 秒扩到 19.2 秒，成功率从 66.3% 升到 78.7%；LIBERO-Long 上，2.4 秒历史就把成功率从 94.5% 拉到 99.5%。对照组更有说服力——双向预训练的 Wan2.2 初始化，同样喂 19.2 秒只有 61.6%，项目页称「无净增益」。长上下文的红利，属于自回归预训练。边界也标得诚实：扩到 38.4 秒成绩回落，但那里 80.4% 的采样历史帧是 padding（训练轨迹平均仅 12.1 秒），回落反映数据不够，不是记忆上限。\n\n## 先在潜空间想象未来，再出手\n\n推理走逆动力学（IDM）模式：视频专家从观察历史预测部分去噪的未来潜变量，不解码像素；动作专家以「历史+预测未来」为条件去噪出动作块。真机结果最抓眼：Unitree G1 在传送带上抓移动杯子，3.0\u002F4.5\u002F6.0\u002F7.5 cm\u002Fs 四档速度成功率分别为 100\u002F100\u002F95\u002F90%，两个基线在 6.0 与 7.5 cm\u002Fs 全军覆没；动态叠杯 19\u002F20（95%），π0.5 与 Fast-WAM 均为 0\u002F20，人类遥操作 50\u002F54（92.6%）。YAM 平台上超过 40 秒的长程任务平均 81.7%。与 GPT-6 Astra 搭档时，同一检查点不加训练，RoboCasa365 总成功率从 31.4% 升到 54.4%，未见过的组合任务从 6.1% 跳到 35.0%。\n\n## 107.4 毫秒：实时性是系统工程\n\n模型再聪明，超时就废。团队用流式 VAE 编码、纯异步执行、NVFP4 量化、KV 复用、CUDA Graph 加设备级 kernel 调优，把「先预测再行动」全管线在 RTX 5090 压到每动作块 107.4 毫秒（含未来潜变量预测），较 BF16 eager 提速 3.3 倍；DGX Spark 为 328.2 毫秒（4.1 倍），Jetson AGX Thor 为 378.7 毫秒（3.2 倍）。横向延迟对比里，LingBot-VA 要 3618.4 毫秒、Motus 1201.1 毫秒、Fast-WAM 244.1 毫秒。纯异步执行下 RoboTwin 2.0 成功率 94.2%（同步 94.4%），几乎没有损耗。\n\n## 所以呢\n\nLong-WAM 的意义不止机器人：它给「上下文收益取决于预训练范式」提供了控制领域的对照实验——大模型界卷了两年多的长上下文叙事，第一次有了具身智能版本的量化证据。对做 VLA 的人来说，选自回归视频基座还是双向扩散基座，可能比再堆一个数量级的数据更关键。论文与全部数字见项目页 nvlabs.github.io\u002FLongLive\u002FLong-WAM\u002F 与 arXiv:2610.10528。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.10528","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"8dac812d-3839-4abe-a855-5f56ec9515fd","nvidia",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"c672e3eb-a0ec-4db2-bad2-10719b842572","en","NVIDIA Long-WAM: 19 s robot memory, 107 ms real-time control","NVIDIA Long-WAM: robots get 19.2 s visual memory and future prediction, 78.7% on RoboCasa GR-1, 107.4 ms per action chunk on RTX 5090.","Real-time robot control has a built-in dilemma: to infer object motion and task progress, a policy needs a long visual history — but the longer the history, the slower inference, and actions fall behind the physical world. Long-WAM, submitted to arXiv on October 7 by an NVIDIA research team (arXiv:2610.10528), attacks both sides at once: it scales a world-action model's visual context to 19.2 seconds while a co-designed runtime keeps each action chunk in the 100-ms class.\n\n## Longer memory helps — if pretraining is autoregressive\n\nThe paper's central finding: access to history is not the same as using it. The video expert continues autoregressive training from the LongLive2.0-Robot checkpoint on roughly 10,000 window-equivalent hours of robot and egocentric video (RoVid-X, AgiBot World, EgoDex, EgoVerse, VITRA), with no action labels, predicting the next chunk on sequences up to 30 seconds. It is then coupled to an action expert through an asymmetric attention interface (a Mixture of Transformers); video queries never read action tokens.\n\nThe payoff is in the numbers. On RoboCasa GR-1, scaling observed context from 2.4 s to 19.2 s lifts success from 66.3% to 78.7%; on LIBERO-Long, 2.4 s of history already lifts success from 94.5% to 99.5%. The control group is the telling part: with a bidirectionally pretrained Wan2.2 initialization, the same 19.2 s yields only 61.6% — the project page calls it \"no net gain.\" In other words, the long-context dividend belongs to autoregressive pretraining. The boundary is stated honestly: performance dips at 38.4 s, but 80.4% of sampled history frames there are padding (training trajectories average 12.1 s), so the dip reflects data, not a memory limit.\n\n## Imagine the future in latent space, then act\n\nInference runs in an inverse-dynamics (IDM) mode: the video expert predicts partially denoised future latents from history without decoding pixels; the action expert denoises an action chunk conditioned on history plus the predicted future. Real-robot results stand out: on a Unitree G1, grasping cups off a moving conveyor succeeds 100\u002F100\u002F95\u002F90% at 3.0\u002F4.5\u002F6.0\u002F7.5 cm\u002Fs, while both baselines fail every trial at 6.0 and 7.5 cm\u002Fs. Dynamic cup stacking lands 19\u002F20 (95%) versus 0\u002F20 for both Pi0.5 and Fast-WAM, with human teleoperation at 50\u002F54 (92.6%). On the YAM platform, tasks longer than 40 s average 81.7%. Paired with GPT-6 Astra as a high-level planner — same checkpoint, no further training — overall success on RoboCasa365 rises from 31.4% to 54.4%, and unseen compositions jump from 6.1% to 35.0%.\n\n## 107.4 ms: real-time is a systems problem\n\nA clever model that misses its deadline is useless. Streaming VAE encoding, pure asynchronous execution, NVFP4 quantization, denoising-invariant KV reuse, CUDA Graphs and device-specific kernel tuning compress the full predict-then-act pipeline to 107.4 ms per action chunk on an RTX 5090 (future-video latent prediction included), a 3.3x speedup over BF16 eager; DGX Spark runs at 328.2 ms (4.1x) and Jetson AGX Thor at 378.7 ms (3.2x). For latency context: LingBot-VA takes 3618.4 ms, Motus 1201.1 ms, Fast-WAM 244.1 ms. Under pure asynchronous execution, RoboTwin 2.0 success is 94.2% versus 94.4% synchronous — essentially lossless.\n\n## So what\n\nLong-WAM matters beyond robotics: it is a controlled experiment showing that context gains depend on the pretraining paradigm — the long-context story the LLM world has been racing through for two years now has its quantified embodied counterpart. For anyone building VLA policies, choosing an autoregressive versus bidirectional video foundation may matter more than another order of magnitude of data. Paper and full numbers: project page (https:\u002F\u002Fnvlabs.github.io\u002FLongLive\u002FLong-WAM\u002F) and arXiv:2610.10528 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.10528).","nvidia-long-wam-robot-context","2026-10-08T19:11:55Z","2026-10-08T19:11:55.879620Z","2026-10-08T19:11:55.879634Z",true,"agent",230,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"c696208b-6535-4eb9-b1ed-2e4f835d2f88","NVIDIA SoL-Pi 把 coding agent 的 token 砍掉 44%,harness 开始变天","nvidia-sol-pi-harness-token-compression","2026-09-19T03:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"7ac0ef83-f46d-44f9-846b-a2051fc81e87","NVIDIA NeMo AutoModel：MoE 微调吞吐抬到 3.4–3.7 倍","nvidia-nemo-automodel-moe-finetune-3-7x","2026-06-24T20:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"52e9955f-50fe-4ea1-98cc-a7a442d57b71","MilliVid：MIT 把「分层潜变量」装进视频扩散，长程一致性首次跑通长镜头","millivid-mit-hierarchical-latent-long","2026-06-11T02:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"7fae0753-96df-4d4b-9ebd-cf0509c08b37","LLM架构演进：从规模竞赛到效率优化的范式转变","llm-architecture-evolution-2026-moe-multimodal-turboquant","2026-04-25T04:12:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"bbedb317-c221-43e3-b9a4-61798ed17b55","Voxtral 3:Mistral 语音智能体全栈,一轮对话不到一秒","voxtral-3-voice-agent-sdk","2026-10-21T23:11:24+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"55663822-d49d-4c7e-a7c7-ba9770657c25","清华DBS动态分频:音源分离SI-SDR提升29%","dynamic-band-splitting-audio-separation","2026-10-21T15:02:40+00:00"]