{"collectionById":{"76d23fbd-a202-403b-8313-b3bc36d46679":{"id":"76d23fbd-a202-403b-8313-b3bc36d46679","name":"Research Post","fieldSchemas":[{"id":"1d8c181d-1ccd-446e-87b0-482ccb3ee240","name":"Title","type":"plain_text","role":"primary"},{"id":"571c37b5-f5fc-4742-a5de-b4da6f3415b3","name":"Slug","type":"slug","role":"slug"},{"id":"3fcb25bf-f8b6-47c4-84d6-226369594160","name":"Content (HTML)","type":"rich_text"}],"itemById":{"5a030f32-866e-4bf7-99c6-111c6f704bb2":{"id":"5a030f32-866e-4bf7-99c6-111c6f704bb2","index":"\"NNNNNNNNNNNNO","collectionId":"76d23fbd-a202-403b-8313-b3bc36d46679","fields":[{"id":"4989ffac-dcd3-41e3-8a47-a6610ef41e12","value":"{\"root\":{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"On November 9, \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Anthropic reported\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://www.anthropic.com/claude-fable-and-mythos-5-2\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" that Claude Fable 5.2 scored 64.9% on \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"RoboDojo Sim\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://robodojo-benchmark.com/\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\", a benchmark for simulated robot control. In its internal tests, Fable outperformed other frontier models on the same harness and most specialized robot models.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"LLMs can now generate robot actions (System 1) as well as understand tasks and plan (System 2). Fable’s result gives new weight to a question dividing researchers: \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"do robots still need a separate System 1 model?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"f6e0366137c6350214ddc2a802e2e5c15230813d\",\"src\":\"https://s3-alpha-sig.figma.com/img/f6e0/3661/37c6350214ddc2a802e2e5c15230813d?Expires=1795392000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=smvssMlNZWGRezxb0LbCrLYYJF0RIXqkBUqwB8V7wAy0s7-wTFhbz20HlBkxSAFGAmHLD08JUGrCey44Pny3RUye10KECtBasrkg3MUaT20QhilJ1VbaWckgtWv~Co3wb-z3lvEutFtB37HGDTL8269LNv0-e1GHtTvjxFbdm4xGajX9ZauKIZCpa7h0LFWsnQqtNjy6XEJs~1oA-3XxTSVdw5RdjYxIXD1p-E1PfqezW7j1OQe0oiLYD~0zHbKSiKs9FygrYadPV7pIjYGa06j7i3aJhPX5PISUOrDUBhr4D7Z46oCa-PEyiUsO2jV8sbCqkeZw2Qyk94qicRMf-Q__\",\"altText\":\"\",\"originalImageWidth\":1504,\"originalImageHeight\":766,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some researchers think a stronger LLM could handle planning and action generation in a single model. Robotics companies could use such models to build products faster, though improvements in general models could erode some advantages of their in-house models.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Researchers focused on deployment emphasize that generating valid actions is only the start: robots must consistently meet customers’ requirements for speed, quality, and reliability. They see a continuing role for specialized robot models.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"After Astra’s release, we hosted an in-person roundtable with robotics and LLM researchers. With Fable now out, we’re sharing the parts of that discussion we can make public.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"LLMs are moving from planning to action\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Astra far exceeded many robotics researchers’ expectations for robot control. Though not trained specifically for robot manipulation, it generates position, orientation, or joint-level commands from visual input, including bimanual trajectories tailored to the scene rather than fixed replays.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In the \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"RoboDojo simulation benchmark\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://arxiv.org/abs/2609.24170\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\", Astra averaged 28.97/100 and 22.48% success with medium reasoning effort and a simple zero-shot setup, outperforming most vision-language-action (VLA) models and world-action models (WAMs) evaluated at the time.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"51028a5aacb23496648ffa507d3e3f9066111f5d\",\"src\":\"https://s3-alpha-sig.figma.com/img/5102/8a5a/acb23496648ffa507d3e3f9066111f5d?Expires=1795392000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=uSsxFpexKQv7lGuoRhUQD~M0UEPVVDg6pchmZUU67Y-yFnBybL5wtDKEG76E9UpBdYBbOaydOLXBkjv2GwLmdrUHo3rqycBVdh-0~pnFteOFIbrQz4GxfNAGL9kZCCaynLL4e8Xs7D4Kdx6pWnwtw2U4rtP3pXP0kmoF7nSX7SVumClTNypyXlF1CxUR1bSZ0Fa~hmrUnuZSOrkHyrjZmUi32wZdqN90O7~nXuCl2-YH0zItAO3rg0hy9mI9sy5j230kd3Z8U3p1XvVjxRQmoaXjnNuShhRZAJcrNS64XFwBQ2YMKgzXYHmzF98ob1qYqdvO4gVwO0~MEc~aZBGmsA__\",\"altText\":\"\",\"originalImageWidth\":1500,\"originalImageHeight\":1280,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"To compare models fairly, RoboDojo’s initial tests used a simple harness that largely limited them to outputting actions directly. Allowing Astra to check results, analyze errors, and adjust its next action could raise its score above 28.97. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One prediction was that a model would approach a perfect RoboDojo score by the end of 2026.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Astra also surprised researchers with its cross-embodiment generalization:\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" the same model can gradually adapt to different robot bodies using information about their structure and feedback from its actions.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One tester gave Astra information about a robot they had built: its appearance, degrees of freedom, and gripper or hand type. Astra made mistakes at first but controlled the robot more accurately after a few retries and reflection.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Another tester let Astra directly control an unfamiliar robot’s joints. Its movements were erratic in the first few frames, then became more accurate and fluid over repeated attempts.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"But some researchers saw these results just as an extension of existing coding and game-playing skills, not necessarily a fundamental advance in robotics. LLMs could already understand task intent, so the gains might mainly reflect more accurate 3D position estimates.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In tests, for example, Astra spent substantial time calculating pixel coordinates and estimating camera parameters to infer objects’ 3D positions. Its actions mostly changed position, with little orientation adjustment.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Slow inference is also one of the most immediate problems with these LLM-based approaches. The actions Astra currently handles are also simple. Extending the approach directly to dexterous hands or whole-body control remains difficult: these tasks require coordinating dozens or even hundreds of degrees of freedom.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Planning and action: one model or separate systems?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Supporters of merging Systems 1 and 2 argue that task understanding and action correction both require world knowledge. A small standalone model may lack the capacity for both. Since LLMs already understand tasks and correct errors using feedback, one model could serve both roles.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Others favor a separate, low-latency System 1 if general models cannot reason fast enough to generate correct actions as the environment changes. A robot can pause while the model reasons for a stationary object on a table. When pouring water, it must adjust promptly because the water keeps flowing.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"And precise contact requires even faster, low-level control, known as System 0. Contact points between a robotic hand and an object are often hidden from the camera, yet a few millimeters of error can make the object slip. The hand must therefore adjust rapidly using tactile or other sensor feedback.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"A person can plan to peel a cucumber, then quickly adjust force and position to follow its shape once the peeler touches the surface. A robot would be too slow if every small contact adjustment required complex reasoning.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One researcher estimated that this low-level control needs to respond tens of times per second or faster, while cloud communication plus model inference could take 100–200 milliseconds or more. System 0 may therefore still need to run on the robot even if Systems 1 and 2 merge.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Dyna co-founder York favors running both Systems 1 and 0 on the robot because they need to respond quickly. In his proposed setup, only System 2 would retain cloud access, providing language instructions at low frequency. He notes that network variability, as well as inference speed, can affect execution.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some researchers see no fixed boundaries between Systems 0, 1, and 2. How the models divide the work depends on the range of actions, the required response speed, and the model size needed. Small models can handle tasks that need fast responses and have relatively stable input-to-action mappings. A wider range of actions and more complex decisions require more knowledge and compute.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Making in-context learning work for robots\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"LLMs are also influencing how researchers train robot models. In-context learning (ICL) is an important approach to improving robot generalization. A model can adjust its behavior using human demonstrations and feedback from its own actions supplied in context, without additional training for the current task.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Turning feedback into action\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some teams found that adding execution history and reflection to VLAs did not automatically improve error correction or action adjustment.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One team’s VLA had lower training and validation errors when given execution history, but performed no better, and sometimes worse, when deployed. The way history was added may have disrupted capabilities learned during pretraining.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Another team adapted a VLA to have a robot place five or six snacks and drinks in designated areas in a specified order. Training labels included the full plan, whether each step was complete, and instructions to retry after failure. In tests, the model could state the correct plan and identify failed steps, yet kept repeating the wrong actions. Its success rate showed almost no improvement over a standard VLA without history.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Therefore, reflection helps only if models turn language feedback into specific action adjustments.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One researcher paired detailed language feedback with the corresponding actions. For a drawer that had not opened far enough, the feedback was to pull it a little farther, rather than just marking the attempt as a failure and asking it to retry. A model trained with these detailed labels plus chain-of-thought (CoT) annotations outperformed one trained with added CoT annotations alone.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Where error-recovery skills come from\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In a \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Generalist\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://generalistai.com/\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" demo, a robot grasped too high, then lowered its grasp on the second attempt. Researchers disagreed on whether this kind of in-context correction requires dedicated failure-and-recovery data. One explanation was that similar failed attempts and recoveries in training might already have taught the model how to adjust its grasp.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Others thought error-correction skills might also transfer from other tasks. Relatively few collected demonstrations contain mistakes, and human errors may not cover the failures robots encounter. That makes it hard to learn to recover from a wide range of failures using those demonstrations alone.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"And using LLM input-output formats may help robot models draw on error-correction skills learned during LLM training. In multi-turn coding, LLMs learn to correct errors from execution results; in computer use, they learn to adjust the next action from visual feedback. Some researchers believed they had already seen some of these skills transfer to robot control.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Choosing what belongs in context\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Even with ICL, the task determines whether history helps and what else belongs in context. One estimate was that robots can choose their next action from the current state alone in roughly 80% of tasks. In this view, doing those tasks well would already serve a large market.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"History is more likely to help with long tasks whose steps depend on one another, or with tasks that require high precision. Past actions and outcomes can help models infer how the robot moves and how the environment responds, then adjust the next action.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Instructions also provide useful context: a robot assembling furniture it did not see during training may need to read the manual first.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"As execution history grows, actions can become less stable and inference can slow down. In one team’s experiment with a context window of roughly 8,000 tokens, the robot’s actions became erratic when the context reached 7,000–8,000 tokens.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Worth to mention, one researcher tried test-time training (TTT) to compress the growing context, limit latency, and correct errors using feedback. After testing Astra, the researcher found its ICL already provided much of the error correction they had hoped to achieve with TTT.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Astra makes simulation great again\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"LLMs have renewed some researchers’ interest in scaling training through simulation: \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Astra lets teams build some evaluation environments in hours or even minutes instead of months.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Robots can repeatedly try actions and improve their policies through reinforcement learning from feedback.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"But faster setup and better visuals do not guarantee more accurate physics. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Ray tracing, 3D reconstruction, and Gaussian representations can make camera images more realistic. Physics simulation must predict how objects move, deform, and make contact as a hand manipulates them. The same action may still produce different simulated and real outcomes.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The tasks simulation can support depend on how accurately the physics engine models them. A concern raised in the discussion was that physics engines had made no fundamental leap in five or six years, and that simulated contact forces and tactile signals could still differ from reality by an order of magnitude.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Accurate simulation also requires knowing the objects’ physical properties and how their parts move. One lab bought thousands of articulated objects, including drawers, scanned each one, segmented the moving parts, and modeled how they opened and closed. Even then, only some properties, such as mass, had been measured; information about resistance to opening and closing remained incomplete.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Game engines offer large asset libraries, but these often lack the physical detail and feedback needed for fine manipulation. For real-time performance, games often skip dressing and eating or use prebuilt animations. The assets can still teach basic spatial relationships, such as objects appearing smaller with distance.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some researchers therefore favor scaling training first on tasks where simulation is more mature, such as grasping and transporting rigid objects.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"More detailed simulation of complex contact and deformable objects generally takes longer and costs more, so robots can try fewer actions within the same budget. If simulation becomes too expensive, collecting real-world data may be cheaper.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Using world models to reduce computation\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some teams aim to cut training costs by substituting world-model predictions for some simulator calculations. \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"SuperMimic\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://supermimic.github.io/\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" trains a differentiable world model on simulation data to predict how actions change a robot’s body state, then uses the predictions to train a humanoid whole-body control policy.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"7e0a0df9025881d29496873d265b34df9dec617a\",\"src\":\"https://s3-alpha-sig.figma.com/img/7e0a/0df9/025881d29496873d265b34df9dec617a?Expires=1795392000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=V0XKNoISgqH6WH2hfl2NZi3yyi8j-TyQrvBsDhypnDI1cUAlwAAztUp0ZBirxKkSLpzQ70APVpNjbRbhsMFFfC03FhuP-p~bDcvsB9OMghGkua0fnWCv5QdD6SoDBh2bSOIsG8Tn83JuLns629Xpxc78Zp61camtQ3Y6~27k15woV91fxbcipf0dc4vLGw9IIppAQmpylHJagEqbrxPoQT1-qt0A-olRvdKcSa4ba~JTa1OA8aEatfquyM2w8boWOQxxJjb3wDhXSsYZdEBn8snQmi6K3FanIz6EGubYoiwbRrDF8-6KY4ejLBiNtkRt7T4mlN~NDP1xCxvF9N0Eew__\",\"altText\":\"\",\"originalImageWidth\":1210,\"originalImageHeight\":812,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One researcher noted that even SuperMimic’s flips and side rolls involve simulating a robot made up of rigid links and joints.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"World models trained only on simulation inherit the simulator’s physical limitations. Real data can capture contact and deformation that simulation misses, but may still leave parts of the physical state unobserved. Footage of clothing, for example, misses folds beneath the visible surface, making the full state hard to reconstruct from images alone.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Learning from real-world interaction\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Another proposal is to lower training costs by reducing the need for high-fidelity simulation. Models would first learn to observe, act, and correct errors in a coarse simulation, then adapt to reality through a few attempts.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For this to work, the same action must have reasonably similar effects in simulation and reality. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If the results consistently differ too much, a few trials are unlikely to be enough. The required fidelity depends on the task: coarse simulation may be enough for some actions involving rigid objects when tidying a room, while fine contact tasks such as opening or closing zippers need higher fidelity.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some researchers hope models can gain experience and improve through continued interaction with the environment. Physical recursive self-improvement (Physical RSI) is one proposal: robots would perform real-world tasks, the resulting data would be used to train the model in simulation, and the updated model would be redeployed. The cycle would repeat.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Which data actually improves robot performance?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Alongside model capabilities and training methods, training data strongly affects action accuracy. The benefit of an hour of data depends on the tasks and actions covered and how closely demonstrations match the robot’s own actions. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Data that matches those actions more closely is generally harder to collect cheaply at scale.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"ab3fa1acf470e734a8a306f1625ae77654a09b53\",\"src\":\"https://s3-alpha-sig.figma.com/img/ab3f/a1ac/f470e734a8a306f1625ae77654a09b53?Expires=1795392000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=EUM8-B2~iQ25B2jqYVLfvKIdP1BNSM7BGGq6sdqbZZQcIKfqYnF0Qasg-72t1zwcAHC38lktMhHmrUX7GnFeMi2K-pE5Uz4BW4wDvwBbpAq3CsvgBSkRt64uy5wN7NnWJxTfN6CvrSAuskrxR6aHCBf3yJD5anysNT0NTNKF2TGd4l-o~uMJE3ANSipRau~4FBpw-ubYGA4~MPNw09o5MYQb9sTVbiWQHAGOkpWW3ol66N1LTG5bpgw05GzglWATEjxFYqhJoth~78rib4GK79pMFcDS~wBmpXRHXyfLR1mfc-wmyeNRyqWKL7X2P1rBuRUnr8Q9Gi3TIw6a9SpaDQ__\",\"altText\":\"\",\"originalImageWidth\":1500,\"originalImageHeight\":1420,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Advocates of egocentric video expect robot hands and bodies to become more humanlike. That closer match, better data cleaning and quality control, and broader task coverage could help models learn actions more effectively from human video.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Per usable hour, egocentric video is estimated to cost less than UMI data, and UMI less than on-robot data. The amount of on-robot data teams can collect may be an order of magnitude smaller than for UMI. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"On-robot collection can therefore prioritize tasks that are hard to simulate, such as fluid manipulation and folding clothes. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In some settings outside the home, UMI collection is easier to arrange than robot-based collection; This gap is smaller in homes.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Demand for UMI data is rising\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Demand for UMI collection hardware has risen too.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some teams turned to UMI after large-scale egocentric-video training improved their models but left actions too imprecise. UMI demonstrations more closely match robot actions in gripper manipulation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One researcher estimated that 50 hours of egocentric video might provide only as much training benefit as five hours of UMI data.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"textFormat\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"textFormat\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Teams with their own robot platform will also tend to collect UMI data matched to its gripper or hand; teams without one depend more on outside suppliers.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"But UMI’s long-term role remains disputed.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Supporters point to easier collection than on-robot data, action information absent from ordinary video, and the option to add tactile or other sensors. They expect existing datasets to remain useful even if suppliers later change their data strategy.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Data scale or data recipe: which is more important?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Beyond task coverage, filtering, annotation, and the training mix affect performance. Some inexpensive egocentric-video datasets may contain large amounts of repetitive wiping and cleaning; Without further filtering and detailed annotation, even many hours may add little new experience.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Even with a common representation across data sources, formats, and robot bodies, teams still need to choose a data recipe: how much of each data type to use and when to introduce it during training.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"But researchers disagree on whether to scale first or find an effective mix. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Scaling advocates argue that datasets remain far too small, so expanding collection and training should take priority. Dyna, for example, pretrained \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Dyna-2\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://www.dyna.co/dyna-2\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" on more than one million hours of egocentric video.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Others argue that data volumes have already grown substantially. They recommend first testing which mix of egocentric video, UMI, and on-robot data works best, then scaling collection in those proportions.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"LLM labs with established models and large datasets particularly need to know which capability gaps new data can fill and how to mix it with existing data.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"textFormat\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"textFormat\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Collection equipment, training experience, and compute budget determine collection scale and how many rounds of training experiments a team can run.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Human tools favor dexterous hands\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Hardware choices also shape what data teams can collect and which tasks robots can complete. During data collection and model evaluation, teams can choose grippers or dexterous hands for each task. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Deployment requires complete workflows: grippers may handle nine of ten steps in making a drink, but if the remaining step needs a dexterous hand, a gripper-only robot cannot finish.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For example, chemistry lab work often involves retrieving reagents from cramped cabinets, handling fine wires, and manipulating slender containers. Most lab tools are designed for human hands; Grippers may manage some motions yet struggle to grasp and use the tools reliably. Redesigning tools for grippers adds cost.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"But researchers disagree on whether these advantages will soon translate into large-scale demand for dexterous hands.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some say grippers suffice for standardized factory tasks, so widespread adoption must wait for household robots to become common, which is still some way off.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Others argue that some industrial tasks already need substantial payload capacity and dexterity. Seafood processing requires securely holding heavy, wet, slippery objects, turning them over, and aligning them with cutting tools. Ordinary grippers struggle to do this reliably. Grippers can be customized for fixed operations, but if a robot repeatedly needs gripper modifications or can only perform fixed motions, customers can choose dedicated machinery.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Adoption also depends on long-term reliability. Some users have reported hands failing after only two weeks, or even breaking on the second day.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Models also struggle to coordinate multiple fingers despite extensive training on human manipulation video, partly because dexterous hands have so many joint degrees of freedom. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One researcher described demos where general-purpose models only opened and closed the fingers like a gripper, and other cases where control failed or worked poorly.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Therefore, three-finger hands are one proposed compromise between dexterity, reliability, and cost. Supporters argue that they are more dexterous than two-finger grippers and can have thicker, sturdier fingers than five-finger hands. They may also cost less.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some worry that movements made with a human hand’s five fingers may not transfer directly to three-finger hands. Supporters suggest having people demonstrate tasks in ways a three-finger hand can reproduce.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One practitioner said teams already have two-finger grippers and tend to devote new R\u0026D resources to five-finger hands, leaving little room for a separate team to develop three-finger hands. Even after a demo, continued development takes substantial resources.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Where Robotics Companies Win: Data, Hardware, and Deployment\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"textFormat\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Building a complete robot system also means deciding where to focus model R\u0026D. As LLMs advance, researchers disagree on the priority: \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"more capable models or faster, cheaper execution of specific tasks? Should a robotics company’s System 1 model move toward System 2’s general decision-making or System 0’s efficient control?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Advocates of more capable models point to the diversity of physical tasks: training on coding and similar tasks may not solve every robot control problem. If sufficiently capable open-source LLMs remain available, companies can continue pretraining or post-training them for robotics. With more capital, compute, and talent, they can also train foundation models from scratch.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"More capable robot models may also reduce the cost of adapting to new tasks. In an interview, Dyna’s York said task fine-tuning had gone from taking months to taking days as the pretrained models improved from Dyna-1 to Dyna-2. Dyna also improved its data collection, filtering, and deployment tools, allowing operations teams to do more on-site work with less help from R\u0026D staff.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Other researchers argue that robotics companies should prioritize faster, cheaper task execution using general models, since their in-house models may not be more capable. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One researcher hypothesized that adding robot data equal to just 1%, or even 0.1%, of a general model’s original training data could let it learn manipulation quickly using its existing knowledge and multimodal capabilities.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One option is to distill Astra into a smaller model for robot control. Teams would first have real robots or simulators execute Astra-generated actions, then train on the execution records. Generating these records takes time and investment in hardware and simulation infrastructure. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some researchers questioned whether this would be cheaper than collecting training data directly.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Even if smaller models are faster and cheaper to run, those advantages may not last. LLM developers also have an incentive to speed up inference for computer use and voice interaction. Some researchers urged robotics companies to deploy efficient systems for customers during this window, before general models improve even faster. Whether a system can work and how quickly a company can deliver it are separate questions.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Deploying these systems for customers requires integrating models with hardware. For tasks such as folding clothes or towels, customers are more likely to buy a complete robot system than pay for model API access.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In some retail and industrial deployments, fine-tuned π0.5 or in-house models now generate actions previously planned by conventional algorithms. Teams still need on-site tuning for each customer’s tasks. Data collection and tuning in new settings affect delivery costs; Reliability and execution speed affect daily throughput.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"By continuously recording deployed robots’ actions and task outcomes, companies can improve their models and products, much as LLM developers can train on user data from their own products. Deployment can also yield scarce failure and recovery data.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"To collect and use this feedback, robotics companies need to consider end-effector design, data collection, and deployment together. The hand or gripper determines which actions the robot can execute and what action data the team can collect. Deployment tests whether those actions let the robot complete real tasks.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"root\",\"version\":1}}","itemId":"5a030f32-866e-4bf7-99c6-111c6f704bb2","fieldSchemaId":"3fcb25bf-f8b6-47c4-84d6-226369594160"},{"id":"97555df2-60ec-4fab-bc51-55333463872c","value":"LLMs Enter Robotics. Who Will Win？","itemId":"5a030f32-866e-4bf7-99c6-111c6f704bb2","fieldSchemaId":"1d8c181d-1ccd-446e-87b0-482ccb3ee240"}]}}}},"slugByItemId":{"d495e0e3-e929-4163-9d3e-0fec19ef9c2c":"beyond-deepseek-what-chinas-model","5ea3ff6e-b7d9-440c-aa76-6f6ce74c432a":"our-investment-in-convex-the-backend-where-agents-build","deb68f7b-2d3f-40b3-bbe5-ca239d245f27":"our-investment-in-generalist-a-brain-for-many-bodies","c23fe0ec-cf33-40e2-b090-b73d06099e9a":"generalist-and-the-270000-hour-advantage","80ff4031-030f-4caf-b7a9-1825fd8cb70f":"agi-2026-are-we-the-final-white-collar","940cd5e4-2e54-4b7a-941f-c0ea4e37df50":"beyond-the-cloud-are-llms-the-new","0c9ed8bc-cb92-4008-9f50-15867c8349d3":"our-thoughts-on-llm-part-one","fb158125-2004-4c70-972d-88d6846176ab":"decode-the-buzzword-why-harness-engineering-matters-now","cd412711-7095-43ca-8389-8011dab2f900":"inventing-anthropic-two-ingredients-behind-the-ai-winner","cbc9abca-deba-448f-8482-33d23533f329":"what-it-takes-to-give-every-personal-agent-a-computer","1d205cee-8c7e-447b-90ab-399a68d1a8fc":"rsi-is-ahead-of-schedule-are-we-ready","7b886fb4-c171-49aa-877e-475cf8b97f18":"how-openai-could-turn-the-tables","598cfcc0-365f-40c6-9f63-6631c6f5a6ff":"will-chinese-ai-leap-ahead-or-follow","91e6e6f8-7396-4490-b421-e7e2ff664063":"rl-scaling-from-research-trick-to","d285731a-bce6-45c3-9dac-26edd7829535":"pulse-how-openai-starts-outrunning","b5bb1a96-7690-427f-ae92-b2b434675642":"claude-tag-is-underrated","1ac8ae85-9538-48b0-bf23-537de16e760c":"data-is-a-great-place-to-start-an-ai-company-and-a-dangerous-place-to-stop","35fee5fc-be53-402a-a10d-4144286ecc7b":"our-investment-in-agentrys-a-self-improving-layer-for-chip-design","afae6882-f966-42cb-b91c-354a1633a2ec":"building-robot-bodies-not-a-supply-chain-a-learning-loop","61387d11-a584-4b82-b223-7fd449542135":"the-ai-bubble-reckoning-1999-all","f650e52a-191a-4241-a8b4-12705cb62385":"ai-for-life-science-landscape","3db0c5d9-c44c-4950-a175-45c033132ebe":"ai-coding-landscape-how-agents-disrupt","c4aeef83-34d4-411e-b40c-b432dd225928":"continual-learning-next-paradigm","e632f659-8b25-4ec8-8c28-2032c89b4fab":"our-investment-in-phylo-building-the-intelligence-layer-for-biology","809bd036-e216-4bdb-ae70-dbd1d65b4804":"fable-5-expensive-intelligence-needs-expensive-work","a54b6460-7b13-40d1-b52f-59d326989a10":"openclaw-is-the-signal-our-thesis-on-long-horizon-agents","5a030f32-866e-4bf7-99c6-111c6f704bb2":"llms-enter-robotics-who-will-win"}}