`rig.ai.policy`

v1.0.0 · stable maturity: draft 5 fields 0 required

Fields

Field	Type		Meaning
`actions`	array<entity>	optional	Optional. `rig.sim.actuator` entities the policy commands
`observations`	array<entity>	optional	Optional. Sensors, bodies, or `rig.anim.track` entities it reads
`task`	string	optional	Optional. Host task name, the same idea as `rig.sim.episode` `task`
`weights`	entity	optional	Optional. `rig.media.asset_ref` of a checkpoint file (`kind` `other`)
`episode`	entity	optional	Optional. `rig.sim.episode` this policy is training or being scored on

A learned map from observations to drives. Format when present. At least one field.

actions and observations are the contract two hosts share. The network, the optimizer, and the replay buffer stay in the host. An unknown checkpoint is kept and not run.

A rig.ai.behavior tree may name an action the host implements by stepping this policy. The tree does not embed the weights.

rig.sim.domain redraws scene fields between episodes. Recorded samples stay on the episode's tracks.

JSON Schema

Resolves at https://rig.works/schemas/rig.ai.policy.schema.json — raw file.

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://rig.works/schemas/rig.ai.policy.schema.json",
  "title": "rig.ai.policy",
  "type": "object",
  "additionalProperties": false,
  "properties": {
    "actions": {
      "type": "array",
      "items": {
        "$ref": "./_defs.schema.json#/$defs/entity"
      }
    },
    "observations": {
      "type": "array",
      "items": {
        "$ref": "./_defs.schema.json#/$defs/entity"
      }
    },
    "task": {
      "$ref": "./_defs.schema.json#/$defs/string"
    },
    "weights": {
      "$ref": "./_defs.schema.json#/$defs/entity"
    },
    "episode": {
      "$ref": "./_defs.schema.json#/$defs/entity"
    }
  },
  "minProperties": 1,
  "x-rig-maturity": "draft"
}

Seen in

robot-lab.json