Manifest Schema Reference

The on-disk manifest is manifest.json. YAML is shown here for readability.

Top Level

yaml

format: policy_package
version: "1.0"
policy: {}
model: {}
hardware: {}

Fields:

Field	Type	Description
`format`	string	Manifest format, usually `policy_package`
`version`	string	Manifest schema version
`policy`	object	Policy identity and source metadata
`model`	object	Artifacts and inference pipeline
`hardware`	object	Expected robot and camera metadata

Policy

yaml

policy:
  name: pi05
  source:
    repo_id: physical-ai/example
    class_path: physicalai.policies.pi05.Pi05

Model

yaml

model:
  n_obs_steps: 1
  artifacts:
    openvino: model.xml
  runner:
    type: action_chunking
    chunk_size: 50
  preprocessors:
    - type: normalize
      artifact: stats.safetensors
  postprocessors:
    - type: denormalize
      artifact: stats.safetensors

Fields:

Field	Type	Description
`n_obs_steps`	integer	Observation history length
`artifacts`	mapping	Logical artifact name to file path
`runner`	`ComponentSpec`	Inference runner
`preprocessors`	list[`ComponentSpec`]	Preprocessing pipeline
`postprocessors`	list[`ComponentSpec`]	Postprocessing pipeline

Hardware

yaml

hardware:
  robots:
    - name: main
      type: SO101
      state:
        shape: [6]
        dtype: float32
        order: [shoulder, elbow, wrist, gripper, lift, rotate]
      action:
        shape: [6]
        dtype: float32
  cameras:
    - name: wrist
      shape: [3, 224, 224]
      dtype: uint8

Robot and camera specs are metadata only. The runtime config still selects the concrete hardware classes.