Datasets:

Voxel51
/

RoboMIND

license: apache-2.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/**
annotations_creators: []
language: en
size_categories:
  - n<1K
task_categories:
  - robotics
pretty_name: RoboMIND (FiftyOne multimodal MCAP subset)
tags:
  - fiftyone
  - multimodal
  - mcap
  - robotics
  - manipulation

RoboMIND → FiftyOne (Native Multimodal MCAP)

A 32-episode subset of x-humanoid-robomind/RoboMIND, eight episodes from each of four robot embodiments (Franka, AgileX, Tien Kung, UR), converted to native multimodal MCAP episodes. Each episode carries per-camera RGB and depth streams, joint telemetry for every recorded arm with timeline plot channels, and the language instruction.

Installation

pip install fiftyone

Usage

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

dataset = fouh.load_from_hub(
    "Voxel51/RoboMIND",
    name="RoboMIND",
    persistent=True,
)
fo.launch_app(dataset)

What you get

32 .mcap episodes across franka, agilex, tienkung, and ur
Per-embodiment telemetry channels discovered from the source layout (for example /puppet-joint-position, /master-joint-velocity-left), each with a timeline plot channel
Per-episode fields: embodiment, task, episode_id, num_frames, duration

License & attribution

The source dataset is released under Apache 2.0. Changes from the source: episode subsetting and conversion from HDF5 to MCAP. The source files carry no per-frame timestamps; frames are timed at the 10 Hz cadence used by RoboMIND's own visualization tooling.

Citation

@article{wu2024robomind,
  title={RoboMIND: Benchmark on Multi-embodiment Intelligence Normative Data for Robot Manipulation},
  author={Wu, Kun and Hou, Chengkai and Liu, Jiaming and Che, Zhengping and others},
  journal={arXiv preprint arXiv:2412.13877},
  year={2024}
}