[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-robottt-scales-robot-memory-8k-timesteps-en":3,"article-related-robottt-scales-robot-memory-8k-timesteps-en":30,"series-research-ae09064b-0052-4d91-94db-8dc8c079efac":77},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":29},"ae09064b-0052-4d91-94db-8dc8c079efac","robottt-scales-robot-memory-8k-timesteps-en","RoboTTT scales robot memory to 8K timesteps","\u003Cp data-speakable=\"summary\">8K timesteps let RoboTTT improve robot control without increasing \u003Ca href=\"\u002Ftag\u002Finference\">inference\u003C\u002Fa> latency.\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>Research org\u003C\u002Fstrong>: Unspecified in arXiv abstract\u003C\u002Fli>\u003Cli>\u003Cstrong>Core data\u003C\u002Fstrong>: 8K timesteps\u003C\u002Fli>\u003Cli>\u003Cstrong>Breakthrough\u003C\u002Fstrong>: Test-time training with fast weights for long-context robot policies\u003C\u002Fli>\u003C\u002Ful>\u003Cp>RoboTTT is trying to solve a practical robot-learning bottleneck: most current robot \u003Ca href=\"\u002Fnews\u002Fdatabricks-query-foundation-models-guide-en\">foundation models\u003C\u002Fa> only work with single-step or short-history visuomotor context. That is fine for simple reactions, but it leaves robots weak on tasks that need memory, adaptation, and multi-stage planning.\u003C\u002Fp>\u003Cp>What makes this paper worth paying attention to is not just that it extends context length. It does so while keeping inference latency flat, which is the part engineers care about when a model moves from a demo to a real robot loop. The paper also claims this longer context unlocks behaviors that short-history policies struggle with, including one-shot imitation from human video, on-the-fly policy improvement, and more robust closed-loop control.\u003C\u002Fp>\u003Ch2>What problem RoboTTT is trying to fix\u003C\u002Fh2>\u003Cp>Robot policies often see only a narrow slice of history. In practice, that means they can miss earlier observations, forget what happened before a perturbation, and fail when a task unfolds over many steps. The abstract frames this as a limitation of recent robot foundation models, which still rely on single-step or short-history visuomotor context.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784269984093-r7j9.png\" alt=\"RoboTTT scales robot memory to 8K timesteps\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>RoboTTT attacks that limitation by treating context length as a scaling axis, not just a nice-to-have feature. The headline number is 8K timesteps of visuomotor context, which the paper says is three orders of magnitude beyond the state of the art policies it compares against. That is a big claim because it suggests robot policies may benefit from the same kind of scaling thinking that has driven progress in other foundation-model settings.\u003C\u002Fp>\u003Cp>For developers, the key issue is not only whether a model can remember more, but whether it can do so without blowing up runtime. A robot policy that needs a much larger inference budget can be hard to deploy in real environments where control loops are tight and latency matters.\u003C\u002Fp>\u003Ch2>How the method works in plain English\u003C\u002Fh2>\u003Cp>RoboTTT stands for Test-Time-Training Robot Policies, and the core idea is to use test-time training inside the policy itself. Instead of treating the model’s state as just activations from a standard recurrent or transformer setup, RoboTTT makes the recurrent state consist of fast weights: parameters that are updated by gradient descent during both training and inference.\u003C\u002Fp>\u003Cp>In plain English, the model does not just read a long history and carry it forward as a hidden vector. It compresses history into weight space, then retrieves that contextual information through those fast weights when making the next action decision. The paper positions this as a way to scale long-context conditioning without increasing inference latency.\u003C\u002Fp>\u003Cp>The method is built on top of robot foundation models such as Vision-Language-Action policies. To make long-context training feasible, the recipe combines sequence action forcing with truncated backpropagation through time. Those are implementation details, but they matter because they show the authors are not just proposing a conceptual trick; they are also addressing the training mechanics needed to make long sequences workable.\u003C\u002Fp>\u003Cp>There is an important distinction here: the paper is not saying every robot should simply memorize more frames. It is saying that if the policy can learn to update fast weights during training and inference, it can carry much longer context in a form that is useful for decision-making.\u003C\u002Fp>\u003Ch2>What the paper actually shows\u003C\u002Fh2>\u003Cp>The strongest concrete result in the abstract is on challenging real-robot manipulation tasks, where RoboTTT improves overall performance by 87% over the single-step context baseline. The abstract also says the system fully completes a five-minute, ten-stage assembly task, which no baseline ever does.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784269983044-fc00.png\" alt=\"RoboTTT scales robot memory to 8K timesteps\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>Another reported result is that RoboTTT trained with 8K-timestep context outperforms the same model pretrained with 1K timesteps by 62%. That matters because it suggests context length itself is acting like a scaling lever, not just a training artifact or a task-specific hack.\u003C\u002Fp>\u003Cp>The paper also claims a first observation: steady gains in closed-loop performance as pretraining context length scales. If that holds up beyond the tasks in the paper, it would be a useful signal for anyone building robot policies that context length should be treated more like compute or data scale, and less like a fixed architectural detail.\u003C\u002Fp>\u003Cp>There are still limits to what the abstract alone tells us. It does not provide the full \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> table, task list, or hardware setup, so you should not read the headline numbers as a universal robotics result. The abstract also does not spell out the exact latency measurements, only that inference latency does not grow with the longer context.\u003C\u002Fp>\u003Ch2>Why engineers should care\u003C\u002Fh2>\u003Cp>If you build robot policies, the main takeaway is that memory may be a first-class scaling dimension. The paper argues that longer visuomotor context can unlock behaviors that short-history models miss, especially on long-horizon manipulation where a task is naturally spread across many stages.\u003C\u002Fp>\u003Cp>That has a few practical implications. First, one-shot imitation from human video demonstrations becomes more plausible when the policy can condition on much longer histories. Second, robustness to perturbations may improve when the model can remember what happened before a disturbance. Third, multi-stage tasks may become tractable without stitching together many brittle sub-policies.\u003C\u002Fp>\u003Cp>The fast-weights approach is also interesting from an engineering perspective because it reframes memory. Instead of storing everything in a larger activation cache, the model learns to update parameters during inference. That is a different design point from standard short-context policies, and it may matter for systems where latency and long-horizon behavior both matter.\u003C\u002Fp>\u003Ch2>What is still unclear\u003C\u002Fh2>\u003Cp>The abstract gives strong headline claims, but it leaves several open questions. It does not say how expensive test-time training is in practice, how stable the fast-weight updates are over long runs, or how sensitive the method is to task type and robot embodiment. Those details matter if you are thinking about deployment rather than just benchmark performance.\u003C\u002Fp>\u003Cp>It is also not clear from the abstract how broadly the 8K-timestep context generalizes across datasets or whether the gains come mainly from the specific combination of sequence action forcing and truncated backpropagation through time. That is the kind of detail practitioners will want before treating RoboTTT as a drop-in upgrade.\u003C\u002Fp>\u003Cp>Still, the paper’s central message is straightforward: robot policies may not be capped by short context forever. If the results hold, context length becomes a new scaling axis for robot foundation models, and that is a meaningful shift for anyone building systems that need memory, adaptation, and long-horizon control.\u003C\u002Fp>\u003Cul>\u003Cli>RoboTTT scales visuomotor context to 8K timesteps without growing inference latency.\u003C\u002Fli>\u003Cli>The model uses test-time training and fast weights to compress history into parameter space.\u003C\u002Fli>\u003Cli>Reported gains include 87% better overall performance and a completed five-minute, ten-stage assembly task.\u003C\u002Fli>\u003C\u002Ful>","RoboTTT pushes robot policy context to 8K timesteps and improves long-horizon control without raising inference latency.","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.15275",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784269984093-r7j9.png","research","en","6312785c-5658-41aa-8eb8-36b0e2d899e0",[17,18,19,20,21],"robot policies","test-time training","fast weights","long-context","vision-language-action",[23,24,25],"8K-timestep context is the paper’s main scaling claim for robot policies.","Test-time training with fast weights is the mechanism used to carry long history.","The abstract reports strong manipulation gains, but leaves deployment details open.",1,"2026-07-17T06:32:30.207954+00:00","2026-07-17T06:32:30.198+00:00","79a71c6c-2a64-4a9d-a7bb-268a8f507489",{"tags":31,"relatedLang":36,"relatedPosts":40},[32,34],{"name":18,"slug":33},"test-time-training",{"name":35,"slug":20},"long context",{"id":15,"slug":37,"title":38,"language":39},"robottt-scales-robot-memory-8k-timesteps-zh","RoboTTT 把機器人記憶拉到 8K","zh",[41,47,53,59,65,71],{"id":42,"slug":43,"title":44,"cover_image":45,"image_url":45,"created_at":46,"category":13},"71d7b8e9-7477-4520-83a0-edad534d32e2","meanflownft-forward-process-rl-average-velocity-en","MeanFlowNFT brings RL to few-step generators","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784271780003-0283.png","2026-07-17T07:02:27.471666+00:00",{"id":48,"slug":49,"title":50,"cover_image":51,"image_url":51,"created_at":52,"category":13},"a586b67c-a662-4cbc-9e3e-193aff904f7c","partition-prompt-aggregate-llm-self-consistency-en","Partition, Prompt, Aggregate: Testing LLM Self-Consistency","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784268177603-pfve.png","2026-07-17T06:02:34.075505+00:00",{"id":54,"slug":55,"title":56,"cover_image":57,"image_url":57,"created_at":58,"category":13},"20f1341a-8779-4c17-a8f0-1274b3ad2de6","metaperch-metadata-bioacoustics-foundation-models-en","MetaPerch uses metadata to boost bioacoustics models","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784185413772-docp.png","2026-07-16T07:02:33.987049+00:00",{"id":60,"slug":61,"title":62,"cover_image":63,"image_url":63,"created_at":64,"category":13},"196a6f70-fce8-4e6f-b6a9-be1f4459541b","ot-ica-wasserstein-linear-ica-en","OT-ICA Uses Wasserstein Distance for Linear ICA","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784183594908-2dd1.png","2026-07-16T06:32:39.335324+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":13},"10ffae7d-4474-45a7-87ee-d7a3f348c5de","mojo-unlabeled-training-neural-decoding-en","MOJO adds unlabeled training to neural decoding","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784181773896-pot3.png","2026-07-16T06:02:25.088864+00:00",{"id":72,"slug":73,"title":74,"cover_image":75,"image_url":75,"created_at":76,"category":13},"a169cb71-ab29-462b-b773-ba2a7dce52aa","claude-j-space-is-not-consciousness-en","Claude's J-space is not consciousness, but it matters","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784100798014-3un1.png","2026-07-15T07:32:39.232247+00:00",[78,83,88,93,98,103,108,113,118,123],{"id":79,"slug":80,"title":81,"created_at":82},"a2715e72-1fe8-41b3-abb1-d0cf1f710189","ai-predictions-2026-big-changes-en","AI Predictions for 2026: Brace for Big Changes","2026-03-26T01:25:07.788356+00:00",{"id":84,"slug":85,"title":86,"created_at":87},"8404bd7b-4c2f-4109-9ec4-baf29d88af2b","ml-papers-of-the-week-github-research-desk-en","ML Papers of the Week Turns GitHub Into a Research Desk","2026-03-27T01:11:39.480259+00:00",{"id":89,"slug":90,"title":91,"created_at":92},"87897a94-8065-4464-a016-1f23e89e17cc","ai-ml-conferences-to-watch-in-2026-en","AI\u002FML Conferences to Watch in 2026","2026-03-27T01:51:54.184108+00:00",{"id":94,"slug":95,"title":96,"created_at":97},"6f1987cf-25f3-47a4-b3e6-db0997695be8","openclaw-agents-manipulated-self-sabotage-en","OpenClaw Agents Can Be Manipulated Into Failure","2026-03-28T03:03:18.899465+00:00",{"id":99,"slug":100,"title":101,"created_at":102},"a53571ad-735a-4178-9f93-cb09b699d99c","vega-driving-language-instructions-en","Vega: Driving with Natural Language Instructions","2026-03-28T14:54:04.698882+00:00",{"id":104,"slug":105,"title":106,"created_at":107},"a34581d6-f36e-46da-88bb-582fb3e7425c","personalizing-autonomous-driving-styles-en","Drive My Way: Personalizing Autonomous Driving Styles","2026-03-28T14:54:26.148181+00:00",{"id":109,"slug":110,"title":111,"created_at":112},"2bc1ad7f-26ce-4f02-9885-803b35fd229d","training-knowledge-bases-writeback-rag-en","Training Knowledge Bases with WriteBack-RAG","2026-03-28T14:54:45.643433+00:00",{"id":114,"slug":115,"title":116,"created_at":117},"71adc507-3c54-4605-bbe2-c966acd6187e","packforcing-long-video-generation-en","PackForcing: Efficient Long-Video Generation Method","2026-03-28T14:55:02.646943+00:00",{"id":119,"slug":120,"title":121,"created_at":122},"675942ef-b9ec-4c5f-a997-381250b6eacb","pixelsmile-facial-expression-editing-en","PixelSmile Framework Enhances Facial Expression Editing","2026-03-28T14:55:20.633463+00:00",{"id":124,"slug":125,"title":126,"created_at":127},"6954fa2b-8b66-4839-884b-e46f89fa1bc3","adaptive-block-scaled-data-types-en","IF4: Smarter 4-Bit Quantization That Adapts to Your Data","2026-03-31T06:00:36.65963+00:00"]