himanshu9series/leisaac-pick-orange-prepared

LeIsaac PickOrange — Prepared Dataset (GR00T-ready)

Pre-processed version of LightwheelAI/leisaac-pick-orange ready for GR00T N1.6 fine-tuning.

What's Different from the Original?

The original dataset requires several modifications to work with GR00T N1.6. This version has all fixes pre-applied:

Fix	Original	This Dataset
Video codec	AV1	H.264 (decord-compatible)
Annotation column	Missing	Integer task_index added to all parquet files
modality.json	Missing	Included (state/action/video/annotation)
Format	LeRobot v2.1	LeRobot v2.1 (compatible)

Dataset Details

• Task: LeIsaac-SO101-PickOrange-v0 (kitchen scene, pick 3 oranges → plate)
• Robot: SO-101 follower (5 arm joints + 1 gripper)
• Episodes: 60 teleoperation demonstrations
• Frames: ~36,000 total
• FPS: 30
• Cameras: Front (480x640) + Wrist (480x640)
• Actions: 6D absolute joint positions in degrees

Action Space

Index | Joint          | Range (deg)
------|----------------|-------------
0     | shoulder_pan   | [-38, 52]
1     | shoulder_lift  | [-100, 64]
2     | elbow_flex     | [-99, 99]
3     | wrist_flex     | [22, 100]
4     | wrist_roll     | [-14, 51]
5     | gripper        | [1, 91]

modality.json

{
    "state": {
        "single_arm": {"start": 0, "end": 5},
        "gripper": {"start": 5, "end": 6}
    },
    "action": {
        "single_arm": {"start": 0, "end": 5},
        "gripper": {"start": 5, "end": 6}
    },
    "video": {
        "front": {"original_key": "observation.images.front"},
        "wrist": {"original_key": "observation.images.wrist"}
    },
    "annotation": {
        "human.task_description": {"original_key": "annotation.human.task_description"}
    }
}

Usage with GR00T

# 1. Clone GR00T
git clone https://github.com/NVIDIA/Isaac-GR00T.git

# 2. Download this dataset
huggingface-cli download rajeshramana/leisaac-pick-orange-prepared \
  --local-dir ./demo_data/pick_orange --repo-type dataset

# 3. Fine-tune
python gr00t/experiment/launch_finetune.py \
  --base-model-path nvidia/GR00T-N1.6-3B \
  --dataset-path ./demo_data/pick_orange \
  --modality-config-path ./so101_pick_orange_config.py \
  --embodiment-tag NEW_EMBODIMENT \
  --num-gpus 1 \
  --max-steps 10000 \
  --no-tune-diffusion-model

Trained Model

A model fine-tuned on this dataset for 10K steps (loss 0.017) is available at: rajeshramana/groot-n1.6-pick-orange

Pre-processing Steps Applied

1. Re-encoded all 120 videos from AV1 → H.264:

   ffmpeg -y -i input.mp4 -c:v libx264 -crf 23 -preset fast -pix_fmt yuv420p output.mp4
   

2. Added annotation column to all 60 episode parquet files:

   df["annotation.human.task_description"] = 0  # integer task_index
   

3. Created modality.json mapping flat columns to GR00T's expected modality structure

Source

• Original dataset: LightwheelAI/leisaac-pick-orange
• Framework: LightwheelAI/leisaac
• Simulation: NVIDIA Isaac Lab 2.3.2

Citation

@misc{leisaac-pick-orange-prepared,
  title={LeIsaac PickOrange Prepared Dataset for GR00T},
  author={Rajesh Kumar},
  year={2026},
  url={https://huggingface.co/datasets/rajeshramana/leisaac-pick-orange-prepared}
}