tshiamor/block-stacking-mimic

Block Stacking MimicGen Dataset

Description

202 demonstrations of a Franka Emika Panda robot stacking colored blocks into a bowl. Generated using NVIDIA MimicGen from source teleoperation demonstrations and augmented with Cosmos world foundation models.

Task

Objective: Pick up 3 colored cubes (Jenga-style blocks) and place them into a bowl on the table.

Success Criteria: All 3 blocks must be placed inside the bowl.

Robot Platform

• Robot: Franka Emika Panda (7-DoF manipulator)
• Gripper: Franka Hand (parallel jaw)
• Simulator: NVIDIA Isaac Lab (Isaac Sim)
• Control Mode: Joint velocity control
• Control Frequency: 30 Hz

Data Modalities

Modality	Key	Shape	Type	Description
Table Camera RGB	obs/table_cam	(T, 200, 200, 3)	uint8	Third-person view
Wrist Camera RGB	obs/wrist_cam	(T, 200, 200, 3)	uint8	Eye-in-hand view
Table Camera Depth	obs/table_cam_depth	(T, 200, 200, 1)	float32	Depth map
Surface Normals	obs/table_cam_normals	(T, 200, 200, 3)	float32	Normal vectors
Segmentation	obs/table_cam_segmentation	(T, 200, 200, 4)	uint8	Instance segmentation
End-Effector Position	obs/eef_pos	(T, 3)	float32	XYZ position
End-Effector Orientation	obs/eef_quat	(T, 4)	float32	Quaternion (x,y,z,w)
Joint Positions	obs/joint_pos	(T, 9)	float32	7 arm + 2 gripper
Joint Velocities	obs/joint_vel	(T, 9)	float32	7 arm + 2 gripper
Gripper Position	obs/gripper_pos	(T, 2)	float32	Finger positions
Cube Positions	obs/cube_positions	(T, 9)	float32	3 cubes x XYZ
Cube Orientations	obs/cube_orientations	(T, 12)	float32	3 cubes x quaternion
Actions	actions	(T, 7)	float32	Joint velocity commands

Dataset Statistics

• Total Demonstrations: 202
• Average Episode Length: ~515 steps
• Total Frames: ~104,000
• File Size: 16 GB

File Format

HDF5 file with the following structure:

cosmos_generated_202.hdf5
└── data/
    ├── demo_0/
    │   ├── actions
    │   ├── obs/
    │   │   ├── table_cam
    │   │   ├── wrist_cam
    │   │   ├── eef_pos
    │   │   └── ...
    │   ├── states/
    │   └── initial_state/
    ├── demo_1/
    └── ...

Usage

Python (h5py)

import h5py
import numpy as np

# Load dataset
with h5py.File('cosmos_generated_202.hdf5', 'r') as f:
    demos = list(f['data'].keys())
    print(f"Number of demos: {len(demos)}")

    # Access a single demo
    demo = f['data/demo_0']
    actions = demo['actions'][:]
    images = demo['obs/table_cam'][:]
    eef_pos = demo['obs/eef_pos'][:]

    print(f"Episode length: {len(actions)}")
    print(f"Image shape: {images.shape}")

PyTorch DataLoader

import h5py
import torch
from torch.utils.data import Dataset, DataLoader

class BlockStackingDataset(Dataset):
    def __init__(self, hdf5_path):
        self.f = h5py.File(hdf5_path, 'r')
        self.demos = list(self.f['data'].keys())

    def __len__(self):
        return len(self.demos)

    def __getitem__(self, idx):
        demo = self.f[f'data/{self.demos[idx]}']
        return {
            'images': torch.from_numpy(demo['obs/table_cam'][:]),
            'actions': torch.from_numpy(demo['actions'][:]),
            'eef_pos': torch.from_numpy(demo['obs/eef_pos'][:])
        }

dataset = BlockStackingDataset('cosmos_generated_202.hdf5')
loader = DataLoader(dataset, batch_size=1, shuffle=True)

Generation Pipeline

1. Source Demonstrations: Human teleoperation via VR hand tracking
2. MimicGen Augmentation: Automated trajectory generation with randomized object poses
3. Cosmos Enhancement: Visual augmentation using NVIDIA Cosmos world foundation models
4. Quality Filtering: 202/714 trajectories passed success criteria (28.3% success rate)

Compatible Frameworks

This dataset can be converted for use with:

• LeRobot (HuggingFace) - Convert to Parquet + MP4
• OpenVLA - Convert to RLDS TFRecord
• Pi0/OpenPI (Physical Intelligence) - LeRobot v2 format
• GROOT N1 (NVIDIA) - LeRobot v2 with 224x224 images
• Cosmos Transfer (NVIDIA) - MP4 videos + JSON annotations

See DATASET_FORMAT_GUIDE.md for conversion instructions.

Citation

If you use this dataset, please cite:

@misc{block_stacking_mimicgen_2026,
  title={Block Stacking MimicGen Dataset},
  author={Tshiamo},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/tshiamor/block-stacking-mimic}
}

License

MIT License

Acknowledgments

• NVIDIA Isaac Lab and MimicGen teams
• NVIDIA Cosmos for visual augmentation
• HuggingFace for dataset hosting