Rooholla/MiM-PushT

MiM Flexiv PushT Dataset

A real-robot dataset collected on a Flexiv Rizon 10S consisting of ~4 hours of (1) random play interactions and (2) push-to-goal episodes, with RGB observations at 5 Hz and 2D delta-position actions.

What’s inside

Tasks / subsets

This repository contains two related subsets:

• Play (random interaction):

  • Unstructured episodes with no explicit start/end goals.
  • Intended for representation learning, world modeling, and offline RL from diverse behavior.

• Push-to-goal (task):

  • Each episode pushes a T-shaped object toward the center of the board.
  • Intended for imitation learning, offline RL, and goal-conditioned learning (if you add goals).

Observation and action spaces

• Observation (images)

  • RGB image at 5 Hz
  • Stored as uint8 in the range [0, 255]
  • Resolution: 256 x 256 x 3
  • Tensor shape per episode in HDF5: 1 x seq_len x 256 x 256 x 3

• Action (actions)

  • 2D delta position command in the board/camera plane: (delta_x, delta_y)
  • Stored as float32
  • Tensor shape per episode in HDF5: 1 x seq_len x 2

Dataset scale (approx.)

• Total recording time: ~4 hours
• Frame rate: 5 Hz
• Approx. total frames: ~72,000 (4 * 3600 * 5)

Note: Exact counts per split are provided in the metadata JSON shipped with each subset.

Data format and files

File layout

.
├── README.md
├── assets/
│ └── preview.gif
├── play/
│ ├── metadata.json
│ ├── shard_000.hdf5
│ ├── shard_001.hdf5
│ └── ...
└── task/
├── metadata.json
├── shard_000.hdf5
├── shard_001.hdf5
└── ...

HDF5 schema

Each HDF5 file contains two main datasets:

• images: shape 1 x seq_len x 256 x 256 x 3, dtype uint8
• actions: shape 1 x seq_len x 2, dtype float32

Metadata JSON schema

Each subset includes a metadata.json describing global stats and shapes. Example:

{
  "num_shards": 13,
  "episodes_per_shard": 1,
  "total_episodes": 13,
  "image_shape":,[256,256,3]
  "action_shape":,[2]
  "min_action": ,[min_1, min_2]
  "max_action": ,[max_1, max_2]
}

How to use

Quickstart (Python)

Below is a minimal example for loading one episode with h5py:

import h5py
import numpy as np

path = "play/shard_000.hdf5"  # or "task/shard_000.hdf5"
with h5py.File(path, "r") as f:
    images = f["images"]   # (seq_len, 256, 256, 3), uint8
    actions = f["actions"] # (seq_len, 2), float32

# Example: normalize images to  float[4]
images_f = images.astype(np.float32) / 255.0

DreamerV4 world model (simulator)

A DreamerV4 world model has been trained on this dataset to enable “simulated” policy training/evaluation inside a learned dynamics model.

World model link: Here

Notes:

• The world model is provided for research convenience and may not perfectly capture real-world contacts/friction.

• Please report issues / inconsistencies via GitHub issues or the Hugging Face discussion tab.

Data collection details

• Platform: Flexiv Rizon 10S (real robot) in-lab setup

• Camera: Orbec Femto Bolt fixed view looking at the board with the T object

• Control/action logged: 2D delta position (x, y) per timestep

• Observation logged: RGB image at 5 Hz

• Play subset: random exploration / interaction

• Task subset: pushing T to board center

Limitations

• Actions are 2D deltas; they do not include full robot state, forces, or contact measurements.

• Camera viewpoint and lab conditions may limit generalization.

License

This dataset is released under AGPL-3.0.

Citation

If you use this dataset, please cite:

text
@dataset{mim_pusht_2025,
  title = {MiM Flexiv PushT Dataset},
  author = {Rooholla Khorrambakht},
  year = {2025},
  url = {https://huggingface.co/datasets/Rooholla/MiM-PushT}
}