spot-gr00t-task0

Spot Robot GR00T Model

This is a fine-tuned GR00T model for Spot robot control, specifically trained for quadruped locomotion tasks.

Model Details

• Base Model: GR00T-N1.5-3B
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Task: Spot robot action generation
• Input Modalities: Video, State, Language
• Output: Robot joint actions (4 legs × 3 joints each)

Usage

from gr00t.model.policy import Gr00tPolicy
from gr00t.data.embodiment_tags import EmbodimentTag

# Load the model
policy = Gr00tPolicy(
    model_path="namaewa-im/spot-gr00t-task0",
    embodiment_tag=EmbodimentTag.NEW_EMBODIMENT,
    device="cuda"
)

# Use for inference
actions = policy.get_action(input_data)

Training Data

The model was fine-tuned on Spot robot demonstration data including:

• Video observations from ego-centric camera
• Joint positions and velocities
• Linear and angular velocities
• Gravity and command information
• Human-annotated task descriptions

Performance

The model achieves competitive performance on Spot robot locomotion tasks with:

• Mean MSE: ~0.55
• Standard deviation: ~0.12

License

This model is released under the MIT License.