{
  "schema_version": "rangoon.article.v1",
  "id": "https://rangoon.ai/insights/archive-physical-intelligence-pi05-april-2025/",
  "url": "https://rangoon.ai/insights/archive-physical-intelligence-pi05-april-2025/",
  "markdown_url": "https://rangoon.ai/insights/archive-physical-intelligence-pi05-april-2025/index.md",
  "json_url": "https://rangoon.ai/insights/archive-physical-intelligence-pi05-april-2025/index.json",
  "title": "π0.5 tested a single VLA on home-cleaning tasks outside its training scenes",
  "summary": "Physical Intelligence introduced π0.5 on April 22, 2025, using mixed robot and multimodal data to test long-horizon cleaning in new homes.",
  "language": "en",
  "author": {
    "name": "Rangoon Editorial",
    "url": "https://rangoon.ai/insights/editorial/"
  },
  "publisher": {
    "name": "Rangoon",
    "maintainer": "Hypler",
    "url": "https://rangoon.ai/"
  },
  "format": "News analysis",
  "topic": {
    "slug": "models",
    "label": "Models",
    "url": "https://rangoon.ai/insights/topics/models/"
  },
  "tags": [
    {
      "slug": "physical-intelligence",
      "label": "Physical Intelligence",
      "url": "https://rangoon.ai/insights/tags/physical-intelligence/"
    },
    {
      "slug": "0-5",
      "label": "π0.5",
      "url": "https://rangoon.ai/insights/tags/0-5/"
    },
    {
      "slug": "robotics",
      "label": "robotics",
      "url": "https://rangoon.ai/insights/tags/robotics/"
    },
    {
      "slug": "vision-language-action",
      "label": "vision-language-action",
      "url": "https://rangoon.ai/insights/tags/vision-language-action/"
    }
  ],
  "dates": {
    "published": "2026-10-23",
    "updated": "2026-10-23",
    "event": "2025-04-22"
  },
  "event_source_id": 1,
  "retrospective": true,
  "reading_minutes": 3,
  "takeaways": [
    "π0.5 separated high-level subtask prediction from continuous low-level action generation inside one model, not a separate human-written planner.",
    "Reported new-home cleaning results were measured on defined homes and subtasks; they do not establish unattended home autonomy.",
    "The launch emphasized generalization over novel settings, while acknowledging mistakes in semantic decisions and motor commands."
  ],
  "source_note": "Historical reporting: Physical Intelligence announced π0.5 on April 22, 2025. The company presented a π0-based VLA co-trained on robot actions, multimodal web data, semantic labels, and verbal instruction demonstrations. Its reported experiments placed mobile manipulators in homes excluded from training data. The event did not announce public weights or a model license, so later OpenPI artifacts are not used as April release proof.",
  "sources": [
    {
      "id": 1,
      "title": "Physical Intelligence: π0.5 launch announcement",
      "url": "https://www.pi.website/blog/pi05",
      "published": "2025-04-22"
    },
    {
      "id": 2,
      "title": "Physical Intelligence: π0.5 technical paper",
      "url": "https://www.pi.website/download/pi05.pdf",
      "published": null
    }
  ],
  "image": {
    "url": "https://rangoon.ai/assets/editorial/archive-physical-intelligence-pi05-april-2025-diagram-v1.webp",
    "thumbnail_url": "https://rangoon.ai/assets/editorial/archive-physical-intelligence-pi05-april-2025-diagram-v1-thumb.webp",
    "alt": "Original research explanatory diagram, π0.5 high-to-low control. Labels: Scene image, Task prompt, Text subtask, Action expert, One-second chunk, Robot joints. Source boundary: dated primary release reporting; original editorial analysis.",
    "type": "original-diagram",
    "caption": "π0.5 high-to-low control. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied.",
    "credit": "Rangoon / Hypler",
    "source_url": "https://rangoon.ai/brand/",
    "rights_url": "https://rangoon.ai/terms/#media-use",
    "rights_note": "Hypler artwork permissions do not grant rights to third-party marks or imply endorsement."
  },
  "sections": [
    {
      "id": "april-event",
      "url": "https://rangoon.ai/insights/archive-physical-intelligence-pi05-april-2025/#april-event",
      "heading": "The April release focused on generalizing across unfamiliar homes",
      "paragraphs": [
        "On April 22, 2025, Physical Intelligence introduced π0.5 for open-world generalization. The company’s experiments put mobile manipulators into kitchens and bedrooms not in training, then asked them to put dishes in a sink, make a bed, or clean a floor. The claim is narrower than a robot that can do any household work: it concerns the team’s behavior on selected cleaning tasks in unseen homes.",
        "The announcement explicitly says the model was far from perfect and did not always succeed on a first attempt. That limit belongs with the videos. A robot must identify relevant objects, choose a suitable subtask, reach without collision, make contact, and monitor progress while the environment changes. A demonstration can show a useful capability under a test protocol without proving dependable operation around people, pets, fragile property, stairs, heat, or unfamiliar objects."
      ],
      "bullets": [],
      "source_ids": [
        1
      ]
    },
    {
      "id": "heterogeneous-training",
      "url": "https://rangoon.ai/insights/archive-physical-intelligence-pi05-april-2025/#heterogeneous-training",
      "heading": "Co-training joined several kinds of supervision",
      "paragraphs": [
        "π0.5 built on π0 but widened its training mixture. Physical Intelligence described robot action data from multiple robots and environments alongside web data, detections, subtask labels, and verbal instructions. The intended effect was to transfer both physical skills and semantic context: a cleaning request needs knowledge of what should be moved, where it belongs, and which intermediate action should come next.",
        "The company’s ablations removed web data, multiple-environment robot data, or cross-embodiment data from the mix. It reported that each source affected its measured follow and success rates, with web data especially helping out-of-distribution object identification in its setup. These are vendor research results rather than an independent ranking of training sources. Their practical implication is still useful: a VLA’s apparent generalization depends on the composition and labeling of its experience, not only its parameter count."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    },
    {
      "id": "two-output-paths",
      "url": "https://rangoon.ai/insights/archive-physical-intelligence-pi05-april-2025/#two-output-paths",
      "heading": "One model produced a textual subtask before an action chunk",
      "paragraphs": [
        "The π0.5 description says the model first predicts a high-level action in text and then turns that subtask into a 50-step, one-second chunk of continuous low-level joint actions. Discrete autoregressive decoding supports the high-level prediction; a flow-matching pathway supplies the motor commands. Calling this a single model does not mean a single output type: language and actions serve different control horizons and are joined during inference.",
        "That architecture can make long tasks easier to inspect because an observer can compare the predicted subtask with the resulting movement. It also introduces two places to fail. A policy can choose an unsuitable goal despite recognizing the scene, or choose a plausible goal and execute its trajectory badly. The launch reports both kinds of mistakes, so a text-like intermediate step should not be treated as a verified plan, a safety check, or a guarantee that the next action is appropriate."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    },
    {
      "id": "evaluation-boundary",
      "url": "https://rangoon.ai/insights/archive-physical-intelligence-pi05-april-2025/#evaluation-boundary",
      "heading": "The test measured subtasks and prompt following in a defined setting",
      "paragraphs": [
        "Physical Intelligence evaluated full cleaning tasks and an out-of-distribution drawer task, measuring average subtask success and language-following rate. It also compared a model trained across varied environments with a baseline that had training data from the test environment. These details matter because a long-horizon headline can conceal how success was scored: partial completion and following the requested instruction are separate observations.",
        "The results are best read as evidence from the company’s homes, objects, robot configuration, prompts, and scoring rubric. The April announcement did not offer public weights or a model license. A later repository or a future checkpoint cannot establish what was available at launch, and neither the demonstrations nor the reported rates establish a safety case for deployment in a home or other unstructured setting."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    }
  ],
  "project_context": null,
  "related_links": [],
  "permissions_url": "https://rangoon.ai/permissions.txt",
  "content_hash": "sha256:f352acbe221735d63015feec7a7730b3d9d2512fbd0c292a6977512f02f65677"
}