Embodiment tags

Explanation of embodiment tags and supported robots

Embodiment tags are used to identify the robot embodiment in your data. They enable GR00T to apply embodiment-specific configurations and handle cross-embodiment training.

Naming convention

Embodiment tags follow the pattern:

<dataset>_<robot_name>

When using multiple datasets for the same robot (e.g., sim GR1 and real GR1), you can drop the dataset name and use only the robot name.

Pretrain embodiment tags

These embodiments were included in the base model pretraining:

ROBOCASA_PANDA_OMRON · string · default: "robocasa_panda_omron"

The RoboCasa Panda robot with omron mobile base.

GR1 · string · default: "gr1"

The Fourier GR1 robot.

Pre-registered posttrain embodiment tags

These embodiments have ready-to-use configurations for fine-tuning:

UNITREE_G1 · string · default: "unitree_g1"

The Unitree G1 robot.

LIBERO_PANDA · string · default: "libero_panda"

The Libero panda robot.

OXE_GOOGLE · string · default: "oxe_google"

The Open-X-Embodiment Google robot.

OXE_WIDOWX · string · default: "oxe_widowx"

The Open-X-Embodiment WidowX robot.

OXE_DROID · string · default: "oxe_droid"

The Open-X-Embodiment DROID robot with relative joint position actions.

BEHAVIOR_R1_PRO · string · default: "behavior_r1_pro"

The Behavior R1 Pro robot.

Custom embodiments

NEW_EMBODIMENT · string · default: "new_embodiment"

Any new embodiment not included in the pre-registered tags.

Use NEW_EMBODIMENT when fine-tuning on your own robot. You’ll need to provide a custom modality configuration.

Using embodiment tags

Embodiment tags are specified in your dataset and during training/inference:

In your dataset

Specify the embodiment tag when creating VLAStepData:

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

step_data = VLAStepData(
    images={"front": [image_array]},
    states={"joint_pos": state_array},
    actions={"joint_pos": action_array},
    text="pick up the cube",
    embodiment=EmbodimentTag.GR1,  # Specify your robot
    is_demonstration=False,
)

During training

Specify the embodiment tag in your training command:

uv run python gr00t/experiment/launch_finetune.py \
    --base-model-path nvidia/GR00T-N1.6-3B \
    --dataset-path <DATASET_PATH> \
    --embodiment-tag UNITREE_G1 \
    --num-gpus 1

During inference

Specify the embodiment tag when loading the policy:

uv run python gr00t/eval/run_gr00t_server.py \
    --embodiment-tag GR1 \
    --model-path nvidia/GR00T-N1.6-3B

Implementation details

Embodiment tags are implemented as an enum in gr00t/data/embodiment_tags.py:14-61:

from enum import Enum

class EmbodimentTag(Enum):
    ##### Pretrain embodiment tags #####
    ROBOCASA_PANDA_OMRON = "robocasa_panda_omron"
    """
    The RoboCasa Panda robot with omron mobile base.
    """

    GR1 = "gr1"
    """
    The Fourier GR1 robot.
    """

    ##### Pre-registered posttrain embodiment tags #####
    UNITREE_G1 = "unitree_g1"
    """
    The Unitree G1 robot.
    """

    # ... additional embodiments

    # New embodiment during post-training
    NEW_EMBODIMENT = "new_embodiment"
    """
    Any new embodiment.
    """

Cross-embodiment training

GR00T’s cross-embodiment architecture allows the model to learn from multiple robot types simultaneously. The embodiment tag is used to:

1. Apply embodiment-specific normalization statistics
2. Load embodiment-specific modality configurations
3. Enable the model to distinguish between different robot morphologies

When fine-tuning on a new embodiment, the model leverages knowledge from all pretrained embodiments, enabling faster adaptation with less data.

Next steps

Modality configs

Configure data processing for your embodiment

Data format

Prepare your data in the correct format

Fine-tuning guide

Fine-tune on your custom embodiment