README.md

fastumi_test

Description

FastUMI Pro robot manipulation dataset in RLDS format.

Task: pick up the object

Dataset Info

• Format: RLDS (Reinforcement Learning Datasets)
• Total Episodes: 12
• Total Steps: 6692
• FPS: 30
• Robot Type: fastumi_pro
• Action Type: absolute
• Image Encoding: JPEG
• Image Shape: (1080, 1920, 3)

Features

Observation

Feature	Shape	Description
observation/image	(1080, 1920, 3)	RGB camera image
observation/state	(7,)	End-effector pose: [x, y, z, roll, pitch, yaw, gripper]

Action

Feature	Shape	Description
action	(7,)	Target pose (next state)

Standard RLDS Fields

Feature	Type	Description
reward	float	1.0 at episode end, 0.0 otherwise
discount	float	0.0 at episode end, 1.0 otherwise
is_first	bool	True for first step
is_last	bool	True for last step
is_terminal	bool	False (demonstrations are successful)
language_instruction	string	Task description

Loading the Dataset

import tensorflow as tf

# Load TFRecords
dataset = tf.data.TFRecordDataset([
    'fastumi_test/1.0.0/fastumi_test-train.tfrecord-00000-of-00001'
])

# Parse function
def parse_episode(serialized):
    features = {
        'episode_id': tf.io.FixedLenFeature([], tf.string),
        'num_steps': tf.io.FixedLenFeature([], tf.int64),
        'steps/observation/image': tf.io.VarLenFeature(tf.string),
        'steps/observation/state': tf.io.VarLenFeature(tf.float32),
        'steps/action': tf.io.VarLenFeature(tf.float32),
        'steps/reward': tf.io.VarLenFeature(tf.float32),
        'steps/is_first': tf.io.VarLenFeature(tf.int64),
        'steps/is_last': tf.io.VarLenFeature(tf.int64),
        'steps/language_instruction': tf.io.VarLenFeature(tf.string),
    }
    return tf.io.parse_single_example(serialized, features)

dataset = dataset.map(parse_episode)

Data Source Mapping

Source Files (per session)

session_YYYYMMDD_HHMMSS/
├── SLAM_Poses/
│   └── slam_raw_baseframe.txt    -> observation/state, action
├── RGB_Images/
│   ├── video.mp4                 -> observation/image
│   └── timestamps.csv            -> temporal alignment
└── Clamp_Data/
    └── clamp_data_tum.txt        (gripper data included in SLAM file)

Citation

If you use this dataset, please cite:

@misc{fastumi_rlds,
    title={FastUMI Pro RLDS Dataset},
    year={2024},
}