Nirav-Madhani/gr00t-g1-palm-pose-augmented

Unitree G1 Fruits Pick & Place — Palm Pose Augmented (FK Traces)

TL;DR. This dataset mirrors NVIDIA’s PhysicalAI-Robotics-GR00T-Teleop-G1 release and adds four forward-kinematics (FK) signals per frame: left_palm_pose, right_palm_pose, and their short-horizon future traces. These compact, goal-centric features make downstream imitation learning and planning/MPC easier, without replaying full videos.

Why this augmented version?

Most real-robot teleoperation sets expose joint space and RGB. For control and policy learning, task intent remains implicit. A 6-DoF palm pose is a strong proxy for near-term manipulation intent (reach, pre-shape, grasp), and a short-horizon future trace provides a goal-aware target that stabilizes behavior cloning / IL pipelines, improves sample efficiency, and helps planners constrain rollouts during contact phases.

• Expose intent explicitly: Poses summarize what the hands are trying to do, not only current joint states.
• Stabilize learning: Conditioning on a near-future target reduces drift versus raw frames/joints alone.
• Speed up planning: Short-horizon palm traces are directly usable as MPC waypoints / trajectory anchors.

These design choices are consistent with findings in Learning from Demonstration and visual foresight / short-horizon planning (see References).

What’s added (over the source dataset)

For every frame we provide:

• left_palm_pose, right_palm_pose — [x, y, z, qw, qx, qy, qz] in robot base frame.
• left_palm_future_trace, right_palm_future_trace — H × 7 matrices (default H=10) of future palm poses.
  Rows beyond episode end are padded with NaN for easy masking.

All original columns from the source dataset are preserved unchanged.

Data fields (additions)

• left_palm_pose / right_palm_pose (float[7]) — 3D position (meters, base frame) + unit quaternion (w, x, y, z).
• left_palm_future_trace / right_palm_future_trace (float[H,7]) — Row t holds the pose at t+1 … t+H. Padded with NaN across episode boundaries.

Note: All other fields (observation.state, action, timestamps, etc.) follow the source dataset schema.

How the palm poses are computed

1. FK model: Unitree MuJoCo model unitree_robots/g1/g1_29dof_with_hand.xml.
2. Frames: Extract wrist and tri-finger tip frames.
3. Palm center: Average of wrist + three fingertip positions + two synthetic wrist-aligned markers (to better capture the palm plane).
4. Orientation: Quaternion from the wrist yaw link (sign-normalized for temporal continuity).
5. Future trace: Concatenate poses from the next H frames (default H=10), pad with NaN once the episode ends.

Caveats: Synthetic markers approximate palm geometry; small biases may exist depending on hand articulation. See Known limitations.

Visualizations

Animated 3D trajectory previews live under media/ (ego-view overlays show temporal evolution of palm markers), with samples for apple, pear, grapes, starfruit.

Reproducibility: regenerate the augmentation

The steps below rebuild the augmented parquet files from the source dataset.

1) Environment

# Python ≥3.10
conda create -n g1-palm python=3.10 -y
conda activate g1-palm

pip install "mujoco>=3.1" "numpy<2" pandas pyarrow scipy
pip install huggingface_hub hf_xet  # optional: faster asset cloning

2) Assets & source

# (A) Unitree MuJoCo assets
git clone https://github.com/unitreerobotics/unitree_mujoco
export UNITREE_MJCF=unitree_mujoco/unitree_robots/g1/g1_29dof_with_hand.xml

# (B) Source dataset (Hugging Face)
huggingface-cli download --repo-type dataset nvidia/PhysicalAI-Robotics-GR00T-Teleop-G1 --local-dir ./source_g1

3) Run augmentation

# Place this script at scripts/augment_gr00t_dataset.py (minimal reference)
python scripts/augment_gr00t_dataset.py ./source_g1 ./augmented_g1 \
  --model-path "$UNITREE_MJCF" --horizon 10

Reference implementation sketch (for scripts/augment_gr00t_dataset.py):

import argparse, numpy as np, pandas as pd, pyarrow.parquet as pq
from pathlib import Path

def process_parquet(in_path, out_path, mjcf_path, horizon):
    # 1) Load table, extract joint positions per frame (e.g., 43D)
    table = pq.read_table(in_path); df = table.to_pandas()
    # 2) FK per frame -> left/right palm pose (7D)
    # 3) Build future traces (H×7), pad with NaN across episode boundaries
    # 4) df['left_palm_pose'] = list-of-7, etc.; df['*_future_trace'] = list-of-lists
    # 5) Save parquet with schema preserved + new columns
    pq.write_table(pd.DataFrame(df).to_arrow(), out_path)

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("src"); ap.add_argument("dst")
    ap.add_argument("--model-path", required=True)
    ap.add_argument("--horizon", type=int, default=10)
    args = ap.parse_args()
    Path(args.dst).mkdir(parents=True, exist_ok=True)
    tasks = ["g1-pick-apple","g1-pick-pear","g1-pick-grapes","g1-pick-starfruit"]
    for task in tasks:
        for pq_in in Path(args.src, task).rglob("*.parquet"):
            rel = pq_in.relative_to(Path(args.src))
            pq_out = Path(args.dst, rel)
            pq_out.parent.mkdir(parents=True, exist_ok=True)
            process_parquet(pq_in, pq_out, args.model_path, args.horizon)

if __name__ == "__main__": 
    main()

4) Quick validation

from datasets import load_dataset
ds = load_dataset("Nirav-Madhani/gr00t-g1-palm-pose-augmented", split="train")

example = ds[0]
for k in ["left_palm_pose","right_palm_pose","left_palm_future_trace","right_palm_future_trace"]:
    assert k in example
    import numpy as np
    arr = np.array(example[k])
    # shape checks
    if "trace" in k: assert arr.shape[1] == 7
    else:            assert arr.shape[0] == 7

Loading & usage examples

from datasets import load_dataset
import numpy as np

ds = load_dataset("Nirav-Madhani/gr00t-g1-palm-pose-augmented", split="train")

# 1) Supervise a goal-aware policy with near-future palm target (e.g., BC)
sample = ds[1234]
obs_joints = np.array(sample["observation.state"])            # (≈43,)
goal_pose  = np.array(sample["right_palm_future_trace"][0])   # +1 step target (7,)

# 2) Evaluate execution drift vs. target trace
trace = np.array(sample["right_palm_future_trace"])           # (H,7) with NaN tail padding
mask  = ~np.isnan(trace).any(axis=1)                          # valid horizon rows

Known limitations

• Palm plane approximation: Two synthetic markers are used to better constrain the palm plane; edge cases may produce slight orientation bias.
• FK model fidelity: Minor kinematic discrepancies can arise between MJCF and the physical hand.
• NaN padding: Consumers should mask NaN rows in future traces near episode ends.

License & attribution

• License: CC BY-4.0 (inherits from source).
• Please cite the source dataset and this augmentation (see below).

Changelog

• v1.0 (2025-09-20): Initial release with four added columns and preview visualizations.

References

• Source dataset: PhysicalAI-Robotics-GR00T-Teleop-G1 (NVIDIA GEAR).
• Robot model: Unitree G1 MuJoCo assets.
• Background: Learning from Demonstration, goal-aware supervision & visual foresight in planning.

BibTeX

@misc{nvidia2025g1teleop,
  title  = {PhysicalAI-Robotics-GR00T-Teleop-G1},
  author = {{NVIDIA GEAR}},
  year   = {2025},
  url    = {https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-Teleop-G1},
  note   = {1000 real-robot teleoperation trajectories on Unitree G1}
}

@misc{madhani2025g1palmaug,
  title  = {Unitree G1 Fruits Pick \& Place — Palm Pose Augmented (FK Traces)},
  author = {Madhani, Nirav},
  year   = {2025},
  url    = {https://huggingface.co/datasets/Nirav-Madhani/gr00t-g1-palm-pose-augmented},
  note   = {Adds palm pose and short-horizon traces for goal-aware IL/MPC}
}

@misc{unitree_mujoco,
  title  = {unitree\_mujoco: MuJoCo models and tools for Unitree robots},
  author = {{Unitree Robotics}},
  year   = {2025},
  url    = {https://github.com/unitreerobotics/unitree_mujoco}
}

@article{argall2009survey,
  title={A survey of robot learning from demonstration},
  author={Argall, Brenna D and Chernova, Sonia and Veloso, Manuela and Browning, Brett},
  journal={Robotics and Autonomous Systems},
  volume={57},
  number={5},
  pages={469--483},
  year={2009},
  publisher={Elsevier}
}

@article{calinon2007learning,
  title={On Learning, Representing, and Generalizing a Task in a Humanoid Robot},
  author={Calinon, Sylvain and Guenter, Franz and Billard, Aude},
  journal={IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics)},
  volume={37},
  number={2},
  pages={286--298},
  year={2007},
  publisher={IEEE}
}

@article{ebert2018visualforesight,
  title   = {Visual Foresight: Model-Based Deep Reinforcement Learning for Vision-Based Robotic Control},
  author  = {Ebert, Frederik and Finn, Chelsea and Dasari, Sudeep and Xie, Annie and Lee, Alex and Levine, Sergey},
  journal = {arXiv:1812.00568},
  year    = {2018}
}
@misc{unifolm-wma-0,
  author = {Unitree},
  title  = {UnifoLM-WMA-0: A World-Model-Action (WMA) Framework under UnifoLM Family},
  year   = {2025},
}

Citation

@dataset{madhani_2025_gr00t_g1_palm_pose_augmented,
  title        = {GR00T-G1 Palm Pose Augmented},
  author       = {Madhani, Nirav},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/Nirav-Madhani/gr00t-g1-palm-pose-augmented},
}