DR_dataset

Simulated robot manipulation dataset for imitation learning.

Dataset Description

This dataset contains 200 episodes of a Trossen WXAI robotic arm performing food transfer tasks in MuJoCo simulation.

Task Description

The robot performs a scooping motion to transfer food from a source container to target bowls:

1. HOME - Start at home position
2. APPROACH_CONTAINER - Move above the source container
3. REACH_CONTAINER - Lower into container
4. SCOOP - Perform scooping motion (wrist rotation)
5. LIFT - Lift from container
6. APPROACH_BOWL - Move above target bowl
7. LOWER_BOWL - Lower to bowl
8. DUMP - Rotate wrist to dump food
9. RETURN - Return to home position

Dataset Structure

ganatrask/DR_dataset/
├── README.md
├── manifest.json
└── data/
    ├── batch_000/
    │   ├── episode_0.hdf5
    │   ├── episode_1.hdf5
    │   ├── ...
    │   └── videos/
    │       ├── episode_0.mp4
    │       ├── episode_1.mp4
    │       └── ...
    ├── batch_001/
    └── ...

HDF5 Episode Format

Each episode_X.hdf5 file contains:

episode_X.hdf5
├── observations/
│   ├── images/
│   │   ├── main_view  (T, 480, 640, 3) uint8 - overhead camera
│   │   └── cam        (T, 480, 640, 3) uint8 - wrist camera
│   ├── qpos           (T, 8) float64 - joint positions
│   └── qvel           (T, 8) float64 - joint velocities
├── action             (T, 8) float64 - joint commands
├── success            bool - episode success flag
├── env_state/
│   ├── source_container  (7,) float64 - [x,y,z,qw,qx,qy,qz]
│   ├── target_container  (7,) float64 - target bowl pose
│   └── bowl_*            (7,) float64 - all bowl poses
└── attrs:
    ├── sim: True
    ├── source: "food_transfer_ik"
    ├── target: bowl name
    ├── dr_enabled: bool
    └── dr_config: JSON (if DR enabled)

MP4 Videos

Each episode has a corresponding MP4 video showing both camera views side-by-side at 50 FPS.

Robot Configuration

• Robot: Trossen WXAI 6-DOF robotic arm
• End-effector: Spoon attached to wrist
• Joints: 6 arm joints + 2 gripper joints
• Action space: 8-dimensional joint position commands
• Cameras: Overhead (main_view) + Wrist-mounted (cam)
• Image resolution: 640x480 RGB

Domain Randomization

This dataset was generated with geometric domain randomization:

Parameter	Value
Position noise	±3.0 cm
Rotation noise	±0.10 rad
Container rotation	±0.15 rad
Bowl count range	1 - 8
Min object spacing	12.0 cm
Container randomization	Yes
90-degree rotation	No

Visual Domain Randomization

This dataset includes visual domain randomization:

Parameter	Value
Table textures	100 variations
Floor textures	100 variations
Container color	Randomized
Bowl color	Fixed
Lighting	Randomized
Light position noise	±0.3 m
Light intensity range	0.5 - 1.2x

Usage

Loading with Python

import h5py
from huggingface_hub import hf_hub_download

# Download a single episode
path = hf_hub_download(
    repo_id="ganatrask/DR_dataset",
    filename="data/batch_000/episode_0.hdf5",
    repo_type="dataset"
)

# Load the episode
with h5py.File(path, "r") as f:
    images = f["/observations/images/main_view"][:]
    actions = f["/action"][:]
    qpos = f["/observations/qpos"][:]
    success = f["success"][()]

print(f"Episode length: {{len(actions)}} timesteps")
print(f"Success: {{success}}")

Loading with datasets library

from datasets import load_dataset

# Load manifest to get episode list
dataset = load_dataset("ganatrask/DR_dataset", split="train")

Citation

If you use this dataset, please cite:

@misc{{{repo_id.split('/')[-1].replace('-', '_')}},
    title={{{repo_id.split('/')[-1]}}},
    author={{Trossen Robotics}},
    year={{{datetime.now().year}}},
    publisher={{HuggingFace}},
    url={{https://huggingface.co/datasets/ganatrask/DR_dataset}}
}}

License

This dataset is released under the MIT License.

Additional Information

• Total batches: 20
• Failed episodes: 0

---

Generated on 2026-01-26 21:06:40 using trossen_arm_mujoco