{
  "schema_version": "rangoon.article.v1",
  "id": "https://rangoon.ai/insights/archive-physical-intelligence-pistar06-recap-november-2025/",
  "url": "https://rangoon.ai/insights/archive-physical-intelligence-pistar06-recap-november-2025/",
  "markdown_url": "https://rangoon.ai/insights/archive-physical-intelligence-pistar06-recap-november-2025/index.md",
  "json_url": "https://rangoon.ai/insights/archive-physical-intelligence-pistar06-recap-november-2025/index.json",
  "title": "Physical Intelligence’s π*0.6 used deployment experience and corrections as VLA training inputs",
  "summary": "Physical Intelligence introduced π*0.6 and its RECAP method on November 17, 2025, combining demonstrations, autonomous rollouts and expert corrections for robot-policy training.",
  "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": "vision-language-action",
      "label": "vision-language-action",
      "url": "https://rangoon.ai/insights/tags/vision-language-action/"
    },
    {
      "slug": "robotics",
      "label": "robotics",
      "url": "https://rangoon.ai/insights/tags/robotics/"
    },
    {
      "slug": "reinforcement-learning",
      "label": "reinforcement learning",
      "url": "https://rangoon.ai/insights/tags/reinforcement-learning/"
    }
  ],
  "dates": {
    "published": "2026-10-24",
    "updated": "2026-10-24",
    "event": "2025-11-17"
  },
  "event_source_id": 1,
  "retrospective": true,
  "reading_minutes": 3,
  "takeaways": [
    "The November event added a training recipe for learning from autonomous rollouts and human corrections, rather than announcing a universal robot controller.",
    "RECAP’s value and advantage signals depend on task outcomes, labels and evaluation design; they do not remove the need to inspect collected experience.",
    "Physical Intelligence’s reported task metrics are supplier results tied to its systems and conditions, not a transferable deployment assurance."
  ],
  "source_note": "Historical reporting: Physical Intelligence announced π*0.6 on November 17, 2025, describing RECAP, a reinforcement-learning method using demonstrations, autonomous experience, outcome feedback and expert interventions. The accompanying paper identifies π*0.6 as an RL-oriented adaptation of π0.6, which it describes as a refinement of π0.5. The throughput and failure-rate figures are Physical Intelligence’s evaluation claims for named tasks and setups; they are not independent validation of a general robot capability, a public model release or a guarantee for another facility.",
  "sources": [
    {
      "id": 1,
      "title": "Physical Intelligence: π*0.6: A VLA that Learns from Experience",
      "url": "https://www.pi.website/blog/pistar06",
      "published": "2025-11-17"
    },
    {
      "id": 2,
      "title": "Physical Intelligence: π*0.6 technical paper",
      "url": "https://www.pi.website/download/pistar06.pdf",
      "published": "2025-11-17"
    }
  ],
  "image": {
    "url": "https://rangoon.ai/assets/editorial/archive-physical-intelligence-pistar06-recap-november-2025-diagram-v1.webp",
    "thumbnail_url": "https://rangoon.ai/assets/editorial/archive-physical-intelligence-pistar06-recap-november-2025-diagram-v1-thumb.webp",
    "alt": "Original robotics explanatory diagram, Experience-driven VLA training loop. Labels: Demonstration, Rollout, Correction, Reward, Value model, Policy. Source boundary: dated primary release reporting; original editorial analysis.",
    "type": "original-diagram",
    "caption": "Experience-driven VLA training loop. 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": "release-event",
      "url": "https://rangoon.ai/insights/archive-physical-intelligence-pistar06-recap-november-2025/#release-event",
      "heading": "The November event centered on an experience-and-corrections training loop",
      "paragraphs": [
        "On November 17, 2025, Physical Intelligence introduced π*0.6 and a method it calls RECAP, short for RL with Experience and Corrections via Advantage-conditioned Policies. The company described a cycle in which a vision-language-action policy begins with demonstrations, collects autonomous episodes on a task, receives outcome feedback and can receive expert corrections, then is trained again with those records. This is a training-method announcement, not a statement that a generic robot can safely learn without supervision. The release makes the data loop explicit: each deployment produces evidence that can improve a later model, but that evidence must be labeled, bounded to its task and evaluated before it becomes a basis for new actions."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    },
    {
      "id": "model-boundary",
      "url": "https://rangoon.ai/insights/archive-physical-intelligence-pistar06-recap-november-2025/#model-boundary",
      "heading": "π*0.6 was presented as an RL adaptation built on a related base model",
      "paragraphs": [
        "Physical Intelligence’s paper says π*0.6 is based on π0.6, and calls π0.6 a refinement of its earlier π0.5 model. It identifies a Gemma 3 4B base VLM and an 860M-parameter action expert, with support for heterogeneous prompts and conditioning information. Those statements explain the announced architecture family, but they do not make the model’s behavior deterministic or establish access conditions for a third party. A model family name can conceal differences in weights, task fine-tuning, input channels and execution hardware. The historical claim is deliberately narrower: the company presented an advantage-conditioned VLA variant designed to use deployment records in reinforcement-learning training. Model availability, license terms and performance on another embodiment require separately dated evidence."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    },
    {
      "id": "experience-data",
      "url": "https://rangoon.ai/insights/archive-physical-intelligence-pistar06-recap-november-2025/#experience-data",
      "heading": "Autonomous rollouts and interventions changed the data provenance that training must retain",
      "paragraphs": [
        "The RECAP description distinguishes demonstrations, autonomous episodes, rewards and teleoperated interventions. That distinction matters for robot-policy governance because these inputs are not interchangeable. A demonstration may encode an intended technique; an autonomous episode may include a near miss or a failure; an intervention records a human correction under a particular state; an outcome label expresses a task definition. Combining them may improve a policy under the vendor’s method, but it also creates a need to retain provenance, task boundaries, sensor context and reasons for intervention. The paper describes a value function that estimates progress and derives an advantage indicator. It does not establish that an advantage estimate is correct for every state, environment change or unseen task."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    },
    {
      "id": "results-boundary",
      "url": "https://rangoon.ai/insights/archive-physical-intelligence-pistar06-recap-november-2025/#results-boundary",
      "heading": "Reported throughput and task runs remained supplier evaluation evidence",
      "paragraphs": [
        "Physical Intelligence reported results for espresso preparation, box assembly and laundry handling, and said RECAP more than doubled throughput on some difficult tasks while reducing failure rates. The paper also reports longer task runs in named environments. These are concrete supplier measurements, more useful than a generic autonomy claim, yet they retain their stated task, hardware, reward and evaluation boundaries. They do not prove that a different robot, home, factory process or maintenance regime will match the same rate. A production decision still needs independently selected success criteria, repeat trials, failure review and operating controls. The November event supports reporting an experience-driven VLA training proposal and its authors’ results; it does not support a claim of independently established general-purpose reliability."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    }
  ],
  "project_context": null,
  "related_links": [],
  "permissions_url": "https://rangoon.ai/permissions.txt",
  "content_hash": "sha256:e97c66f55b2c40e875f2dfc0706d123c9d5b592c17fdb08301f0d6ba6d0a1730"
}