Skip to main content
Embodiment tags are used to identify the robot embodiment in the data. They enable cross-embodiment training by providing consistent identifiers across datasets.

​
Naming convention

Embodiment tags follow the pattern <dataset>_<robot_name>. If using multiple datasets with the same robot (e.g., sim GR1 and real GR1), you can drop the dataset name and use only the robot name.

​
EmbodimentTag

from gr00t.data.embodiment_tags import EmbodimentTag
Enum class for robot embodiment identifiers.

​
Pretrain embodiment tags

These embodiments are used during the pretraining phase:

​
ROBOCASA_PANDA_OMRON

The RoboCasa Panda robot with Omron mobile base. 
EmbodimentTag.ROBOCASA_PANDA_OMRON
# Value: "robocasa_panda_omron"

​
GR1

The Fourier GR1 robot. 
EmbodimentTag.GR1
# Value: "gr1"

​
Posttrain embodiment tags

These embodiments are pre-registered for post-training:

​
UNITREE_G1

The Unitree G1 robot. 
EmbodimentTag.UNITREE_G1
# Value: "unitree_g1"

​
LIBERO_PANDA

The Libero Panda robot. 
EmbodimentTag.LIBERO_PANDA
# Value: "libero_panda"

​
OXE_GOOGLE

The Open-X-Embodiment Google robot. 
EmbodimentTag.OXE_GOOGLE
# Value: "oxe_google"

​
OXE_WIDOWX

The Open-X-Embodiment WidowX robot. 
EmbodimentTag.OXE_WIDOWX
# Value: "oxe_widowx"

​
OXE_DROID

The Open-X-Embodiment DROID robot with relative joint position actions. 
EmbodimentTag.OXE_DROID
# Value: "oxe_droid"

​
BEHAVIOR_R1_PRO

The Behavior R1 Pro robot. 
EmbodimentTag.BEHAVIOR_R1_PRO
# Value: "behavior_r1_pro"

​
Custom embodiments

​
NEW_EMBODIMENT

Placeholder tag for any new embodiment during post-training. 
EmbodimentTag.NEW_EMBODIMENT
# Value: "new_embodiment"

​
Usage

from gr00t.data.embodiment_tags import EmbodimentTag
from gr00t.data.types import VLAStepData

# Create VLA step data with embodiment tag
vla_step = VLAStepData(
    images={"front_cam": [image_array]},
    states={"joint_positions": state_array},
    actions={"joint_velocities": action_array},
    text="Pick up the apple",
    embodiment=EmbodimentTag.UNITREE_G1,
)

# Access embodiment value
print(vla_step.embodiment.value)  # "unitree_g1"

Build docs developers (and LLMs) love

Get started for freeTalk to us