README.md

metadata

language:
  - en
  - zh
tags:
  - robotics
  - manipulation
  - trajectory-data
  - multimodal
  - embodied-ai
multimodal: vision+language+action
license: other
task_categories:
  - robotics
dataset_info:
  features:
    - name: rgb_images
      dtype: image
      description: Multi-view RGB images
    - name: slam_poses
      sequence: float32
      description: SLAM pose trajectories
    - name: vive_poses
      sequence: float32
      description: Vive tracking system poses
    - name: point_clouds
      sequence: float32
      description: Time-of-Flight point cloud data
    - name: clamp_data
      sequence: float32
      description: Clamp sensor readings
    - name: merged_trajectory
      sequence: float32
      description: Fused trajectory data
  configs:
    - config_name: default
      data_files: '**/*'

FastUMI Pro – Multimodal Sample Dataset

Small-Scale Demonstration Data from the FastUMI Pro Multimodal Sensing System
(Only Dozens of Trajectories — Full Dataset Available Upon Request)

Project Homepage

---

📖 Overview

The FastUMI Pro Sample Dataset provides a public preview of the multimodal sensing capabilities of the FastUMI Pro data collection system.

This release contains only dozens of sample trajectories and is intended for:

• System testing
• Robotics and AI pipeline integration
• Preliminary algorithm development
• Demonstrating multimodal alignment and synchronization

Full-scale datasets are available upon request for research or enterprise collaboration.

---

📊 Data Specifications

Purpose: The following table is to provide users with an overview of the technical specifications of the dataset.

Data Type	Path	Shape	Type	Description
RGB Images	RGB_Images/Frames/*.mp4	(H, W, 3)	uint8	Multi-view RGB images
ToF PointClouds	ToF_PointClouds/PointClouds/*.pcd	variable	pcd	Time-of-Flight point clouds
Clamp Data	Clamp_Data/clamp_data_tum.txt	(N, 2)	float	Timestamp + clamp width
Merged Trajectory	Merged_Trajectory/merged_trajectory.txt	(N, 8)	float	Fused multi-sensor pose

---

🧭 Data Formats

All pose data (SLAM, Vive, fused) follow the same structure:

timestamp  x  y  z  qx  qy  qz  qw

Field	Description	Field	Description
timestamp	Unix timestamp	qx	Quaternion X component
x	Position X (meters)	qy	Quaternion Y component
y	Position Y (meters)	qz	Quaternion Z component
z	Position Z (meters)	qw	Quaternion W component

---

Coordinate System

To ensure correct visualization and control, all pose data adheres to the following right-handed coordinate system (World Frame).

• Origin (0,0,0): Geometric center of the tracking base stations (World Frame).
• 🔴 X-Axis: Points Forward (the primary direction of manipulation).
• 🟢 Y-Axis: Points Right (relative to the workspace).
• 🔵 Z-Axis: Points Upward (opposite to the direction of gravity).
  
  Visual reference for the coordinate system.
  Tip: When using simulation environments like ROS or Isaac Gym, ensure your coordinate frame conventions match. You may need to apply a transformation if your framework uses a different "up" axis (e.g., Z-up vs. Y-up).

---

📸 How We Collect Data

We collect data using the FastUMI Pro hardware suite. This system integrates high-frequency sensors to capture comprehensive multimodal interaction data:

• Visual: Industrial-grade RGB cameras.
• Spatial: Time-of-Flight depth sensors for dense 3D reconstruction.
• Haptic/State: Force-sensitive clamp sensors for precise gripper feedback.

---

📥 Download

huggingface-cli download FastUMIPro/example_data_fastumi_pro_raw \
  --repo-type dataset \
  --local-dir ./fastumi_sample/

Optional:

export HF_ENDPOINT=https://hf-mirror.com

---

⚠️ Dataset Scale Notice

This dataset contains only a small number of sample episodes and is not intended for large-scale training.

For full multimodal datasets or enterprise collaborations, please contact the FastUMI team.

---

📞 Contact

Lead: Ding Yan
Email: dingyan@lumosbot.tech
WeChat: Duke_dingyan

---