{
  "docId": "019dd923-5e88-73ef-bd5d-06b04d219fea",
  "docSlug": "dd91c78f6570bf29",
  "documentTitle": "2023 Air Street Capital The State of AI Report 2023",
  "authorId": "AirStreetCapital",
  "authorName": "Air Street Capital",
  "documentKindSlug": "consulting-deck",
  "documentKindLabel": "Consulting deck",
  "sourceTypeSlug": "vc_research",
  "sourceTypeLabel": "VC research",
  "presentationDate": null,
  "orientation": "landscape",
  "aspectRatio": 1.777,
  "pageNumber": 46,
  "pageCount": 163,
  "prevPage": 45,
  "nextPage": 47,
  "slideType": "client_example",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 6,
  "notes": "The slide uses a circular process diagram to explain the self-improving training loop of the RoboCat agent.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/46",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-46",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "RoboCat is a foundation agent for robotic manipulation that can generalise to new tasks and new robots in zero-shot or few-shot (100-1000 examples).",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a224-8b9641d07bb9",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.5,
        "x": 0.25,
        "y": 0.68
      },
      "kind": "diagram",
      "text": "Circular training loop diagram showing data collection, fine-tuning, and self-generation of training data for RoboCat.",
      "attrs": null,
      "subkind": "process",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b21f1787-846d-40fb-8bff-097eea31e8cc",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.95,
        "x": 0.02,
        "y": 0.31
      },
      "kind": "list",
      "text": "It's built on top of DeepMind's multi-modal, multi-task and multi-embodiment Gato. It uses a frozen VQ-GAN tokenizer trained on a variety of vision and control datasets. While Gato only predicted actions, RoboCat additionally predicts future VQ-GAN tokens.\nIn terms of policy learning, the paper only mentions behaviour cloning. RoboCat is fine-tuned with few demonstrations (via teleoperation) and re-deployed to generated new data for a given task, self-improving in subsequent training iterations.\nRoboCat can operate 36 real robots with different action specifications, in 253 tasks on 134 real objects at an impressive speed (20Hz).",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "61f714f3-e9a3-46f6-99c3-ec6081ec24dd",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "real-time performance: 20Hz",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a224-8c65ec1850c8",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.03,
        "w": 0.6,
        "x": 0.02,
        "y": 0.02
      },
      "kind": "other",
      "text": "Introduction | Research | Industry | Politics | Safety | Predictions",
      "attrs": null,
      "subkind": "unclassified",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "73d027b2-da3c-4137-9696-d2e18108de1f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "RoboCat is a foundation agent for robotic manipulation that can generalise to new tasks and new robots in zero-shot or few-shot (100-1000 examples). Impressive real-time performance on a variety of platforms.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9e9071c9-0aff-4d10-b46c-f35414ac0ccf",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.6,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "From vision-language models to low-level robot control: RoboCat",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "feaa47ec-0b12-4c81-a15d-1f9f58c711f3",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 120,
      "from": 11,
      "beatId": "019dd95a-0682-776c-8e35-41afd44ef59f",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Research + Industry sections inventory model, compute, funding facts.",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 120,
      "from": 11,
      "beatId": "019dd95a-0682-776c-8e35-523bfb7f96e6",
      "arcName": "The Mountain",
      "arcSlug": "mountain",
      "beatName": "Rising Action",
      "beatSlug": "mountain-rising-action",
      "evidence": "Escalating capabilities, compute concentration and capital flows.",
      "position": 2,
      "confidence": 45,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [
    {
      "to": 49,
      "from": 41,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3d-f0e6f3668fb1",
      "evidence": "VLM, autonomous driving, PaLM-E, RT-2, RoboCat, drone-racing, CICERO stack as parallel evidence.",
      "position": 5,
      "objective": "Multimodal models extend from vision-language to embodied robots",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 75,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}