ACT PushT — external metadata assessment

Evidence boundary

The validator inspected the listed metadata files only. It did not download weights, import policy code, run inference, control hardware, execute simulation or verify model-card results.

{
  "method": "metadata_only",
  "assessor": "Known Robot",
  "validator_version": "1.3.0",
  "assessed_at": "2026-09-24T15:30:00.000Z",
  "executed_policy_code": false,
  "evaluated_policy": false,
  "established_compatibility": false,
  "status": "incomplete"
}

Generated portable manifest

This is the validator output, including nulls. The separately pinned Hub revision above establishes this assessment snapshot; it is not rewritten into the raw manifest as if the validator discovered it.

{
  "schema_version": "1.0",
  "skill": {
    "name": "Aadarshram Act Pusht",
    "version": null,
    "source": {
      "type": "local",
      "repository": null,
      "revision": null
    }
  },
  "policy": {
    "framework": "lerobot",
    "framework_version": null,
    "architecture": "act"
  },
  "hardware": {
    "robot_family": null,
    "gripper": null,
    "sensors": []
  },
  "runtime": {
    "control_frequency_hz": null,
    "observation_shape": {
      "observation.image": {
        "type": "VISUAL",
        "shape": [
          3,
          96,
          96
        ]
      },
      "observation.state": {
        "type": "STATE",
        "shape": [
          2
        ]
      }
    },
    "action_shape": {
      "action": {
        "type": "ACTION",
        "shape": [
          2
        ]
      }
    },
    "dependencies": []
  },
  "dataset": {
    "repository": "lerobot/pusht",
    "revision": null,
    "schema": null
  },
  "compatibility": [],
  "evaluations": []
}

Missing-information report

Missing declarations remain visible. Model-card prose does not silently satisfy portable metadata fields.

• skill.version — Artifact version is missing.
• skill.source.repository — Source repository is missing from inspected metadata.
• skill.source.revision — Immutable source revision is missing from inspected metadata.
• policy.framework_version — Framework version is not pinned.
• hardware.robot_family — Robot family is not declared.
• hardware.gripper — Gripper is not declared.
• hardware.sensors — Sensors are not declared.
• runtime.control_frequency_hz — Control frequency is not declared.
• runtime.dependencies — Runtime dependencies are not declared.
• dataset.revision — Dataset revision is not pinned.
• dataset.schema — Dataset schema is not declared.
• dependency.unresolved — Resolved dependency evidence is missing.

Cautions

• evaluations — No measured evaluation was found.
• compatibility — No compatibility evidence was found.

Upstream model-card claims

These are attributed summaries, not Known Robot measurements or endorsements.

• The model card says the ACT policy was trained on the PushT dataset for 80,000 steps. Inspect the pinned model card ↗

Inspected file inventory

Hashes make the metadata snapshot auditable without republishing upstream checkpoint weights.

[
  {
    "path": "README.md",
    "sha256": "33142fae10336df5c9137f3da809476ecb821c871636b814086f3a23df9bef89",
    "bytes": 1953
  },
  {
    "path": "config.json",
    "sha256": "541ce265150b3f45ee7c8d43d308facfc31a8e07f8d8bcb95bb5fc3df86bb981",
    "bytes": 1478
  },
  {
    "path": "train_config.json",
    "sha256": "eab635d09399d34cdcd668942f3de5c31a7103eea0fbe6d296902e03cabbd4a6",
    "bytes": 4599
  },
  {
    "path": "policy_preprocessor.json",
    "sha256": "39af97ab0d52b54f441fce4925a27b97fd2c20c67f17099d0c63c4746d5169f5",
    "bytes": 1145
  },
  {
    "path": "policy_postprocessor.json",
    "sha256": "9ca3b5a1986e8b6d4f41238446eb7b3637776853102ae7f9e8ba29700e6a2e66",
    "bytes": 660
  }
]

Limitations and next evidence

• Only repository metadata was inspected; weights were not downloaded and policy code was not executed.
• The pinned Hub revision is assessment provenance and was not inferred by the validator from the local metadata snapshot.
• No success rate, safety claim, physical transfer claim or compatibility result is established.

A future measured evaluation must be published as a separate evaluation record with an execution protocol, trials, outcomes and evidence. It will not overwrite or retroactively strengthen this assessment.