Add A Robot

A robot defines concrete enum vocabularies and binds semantic robot roles to those members.

from retarget import (
    RobotGeometry,
    RobotJoint,
    RobotKind,
    RobotLink,
    RobotRole,
)
from retarget.robots import RobotSpec, RobotVocabulary, robots


class LabRobot(RobotKind):
    BIPED = "lab_biped"


class LabRole(RobotRole):
    PELVIS = "pelvis"
    LEFT_FOOT = "left_foot"


class LabJoint(RobotJoint):
    HIP = "hip"
    KNEE = "knee"


class LabLink(RobotLink):
    PELVIS = "pelvis_link"
    LEFT_FOOT = "left_foot_link"


class LabGeometry(RobotGeometry):
    LEFT_FOOT = "left_foot_geom"


@robots.register(LabRobot.BIPED)
def lab_robot() -> RobotSpec[LabJoint, LabLink, LabGeometry, LabRole]:
    return RobotSpec(
        name="lab_biped",
        height_m=1.0,
        vocabulary=RobotVocabulary(
            joints=LabJoint,
            links=LabLink,
            geometries=LabGeometry,
            roles=LabRole,
        ),
        joints=tuple(LabJoint),
        links=tuple(LabLink),
        geometries=tuple(LabGeometry),
        contact_links=(LabLink.LEFT_FOOT,),
        joint_limits={
            LabJoint.HIP: (-1.0, 1.0),
            LabJoint.KNEE: (-2.0, 0.0),
        },
        joint_roles={LabRole.PELVIS: LabJoint.HIP},
        link_roles={
            LabRole.PELVIS: LabLink.PELVIS,
            LabRole.LEFT_FOOT: LabLink.LEFT_FOOT,
        },
    )

RobotSpec never normalizes these members to strings. Passing a member from a different enum class fails even if its value matches.

Nominal tracking selections belong in an optimization profile, not the robot spec. MuJoCo body aliases remain strings because they are foreign model identifiers.

File Specs

Serialized robot specs reference importable enum classes explicitly:

name = "three_point_bot"
height_m = 1.0
joints = ["root_sway", "left_leg", "right_leg"]
links = ["pelvis", "left_foot", "right_foot"]
contact_links = ["left_foot", "right_foot"]
geometries = []

[vocabulary]
joints = "examples.vocabularies:DemoRobotJoint"
links = "examples.vocabularies:DemoRobotLink"
geometries = "examples.vocabularies:DemoRobotGeometry"
roles = "examples.vocabularies:DemoRobotRole"

The strings are deserialized into members of those exact classes. See examples/custom_robot.toml for the complete file and examples/custom_robot.py for the programmatic equivalent.

from retarget.robots import RobotSpec

robot = RobotSpec.load("examples/custom_robot.toml")