BlockStackingSpecifiedOrderTask/README.md

# RoboLab Motion-Planning GR00T Dataset

This dataset was generated from RoboLab/Isaac Sim scripted Cartesian motion planning.
It is stored in the GR00T-flavored LeRobot v2 layout expected by NVIDIA Isaac-GR00T.

## Contents

- Episodes: 80
- Frames: 61839
- FPS: 15
- Robot type metadata: `droid_abs_ik`
- State/action dimension: 8

## Tasks

- `0`: Stack the blocks in the order from bottom to top: red, blue, green, yellow

## Directory Layout

```text
meta/info.json
meta/modality.json
meta/tasks.jsonl
meta/episodes.jsonl
meta/stats.json
meta/relative_stats.json
data/chunk-000/episode_*.parquet
videos/chunk-000/<video_key>/episode_*.mp4
robolab_motionplanning_config.py
README.md
```

## Data Format

Each parquet row contains:

- `observation.state`: float32 list `[x, y, z, qw, qx, qy, qz, gripper]`
- `action`: float32 list with the next end-effector target in the same format
- `timestamp`: seconds at 15 FPS
- `annotation.human.action.task_description`: integer index into `meta/tasks.jsonl`
- `task_index`, `episode_index`, `index`
- `next.reward`, `next.done`

`meta/modality.json` splits state and action into:

- `eef_position`: indices `[0:3]`
- `eef_quaternion_wxyz`: indices `[3:7]`
- `gripper`: indices `[7:8]`

Video modalities:

- `observation.images.front`
- `observation.images.wrist`

## Fine-Tuning GR00T

Clone and install Isaac-GR00T following the official repository instructions:

```bash
git clone https://github.com/NVIDIA/Isaac-GR00T.git
cd Isaac-GR00T
```

Then fine-tune with this dataset path and the included modality config:

```bash
export NUM_GPUS=1
CUDA_VISIBLE_DEVICES=0 uv run python   gr00t/experiment/launch_finetune.py   --base-model-path nvidia/GR00T-N1.7-3B   --dataset-path /path/to/this/dataset   --embodiment-tag NEW_EMBODIMENT   --modality-config-path /path/to/this/dataset/robolab_motionplanning_config.py   --num-gpus $NUM_GPUS   --output-dir /tmp/robolab_motionplanning_gr00t   --save-total-limit 5   --save-steps 1000   --max-steps 2000   --global-batch-size 32   --dataloader-num-workers 4
```

For open-loop evaluation after training:

```bash
uv run python gr00t/eval/open_loop_eval.py   --dataset-path /path/to/this/dataset   --embodiment-tag NEW_EMBODIMENT   --model-path /tmp/robolab_motionplanning_gr00t/checkpoint-2000   --traj-ids 0   --action-horizon 16   --steps 260   --modality-keys eef_position eef_quaternion_wxyz gripper
```

Notes:

- This dataset stores Cartesian end-effector pose with quaternion orientation.
- The included GR00T config treats the three action slices as `NON_EEF` absolute vectors.
- For a production EEF-specific setup, convert quaternion orientation to a GR00T-supported EEF rotation format such as 6D rotation, then use the corresponding GR00T `ActionType.EEF` / action format.
- If you change the action horizon or modality config, regenerate GR00T statistics as described in the Isaac-GR00T data config guide.

References:

- https://github.com/NVIDIA/Isaac-GR00T
- https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/data_preparation.md
- https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/finetune_new_embodiment.md
- https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/data_config.md