RealMan said it hopes to break data silos and accelerate embodied intelligence research with RealSource. | Source: RealMan Robotics
RealMan Intelligent Technology Co. announced the open-source release of RealSource, its high-quality, multi-modal robot dataset. The company said it designed this dataset to address the industry’s shortage of fully aligned real-world data.
The dataset is built entirely on 10 real-world simulated environments within the company‘s Beijing Humanoid Robot Data Training Center. Opened in August, this training center brings together core technology R&D, scenario-based application testing, operator training, and ecosystem collaboration.
When creating the dataset, RealMan said it focused on data quality and complete multi-modal coverage. Founded in 2018, the Beijing-based company creates robotic arms and mobile robots that cater to retail, food service, commercial services, inspections, healthcare, education, aerospace, and industrial production.
RealSource covers 10 real-world scenarios
RealMan Robotics built the dataset at its 3,000 m² (32,291.7 sq. ft.) Beijing Humanoid Robot Data Training Center, which includes:
- Training Zone: This provides high-volume, efficient robot training for foundational manipulation tasks.
- Scenario Zone: Ten real-world environments in this “Robot University” include smart home and eldercare, daily living, agriculture, new retail, automotive assembly, and catering.
Robots perform tasks such as opening refrigerator doors, folding laundry, and sorting materials on factory lines, capturing data in realistic, noisy, and diverse environments. Data collection is conducted outside the “laboratory greenhouse,” directly addressing the complexity of daily life, said RealMan.
This ensures high realism, strong practicality, and superior generalization across scenarios, the company claimed. Key metrics include 100% modality completeness, 78% noise resistance, and 82.1% smoothness.
The team used three robots for data collection. The first is the RS-01, a wheeled folding mobile robot with 20 degrees of freedom (DoF) and multi-modal vision.
The second is RS-02, a dual-arm lifting robot with RGB and depth vision, dual 7-DoF arms, 9 kg (19.8 lb.) payload per arm, six-axis force sensing, and overhead fisheye perception. The third is RS-03, a dual-arm, dual-eyed robot with a binocular system for high-resolution stereo vision and precise manipulation.
All three robots integrate large field of view (FOV) wrist and head cameras (H 90° / V 65°) and full spatiotemporal synchronization, according to RealMan.
RealMan touts advantages of multi-modal data
The RealSource dataset covers the full perception-decision-execution chain, integrating RGB images, joint angles and velocities, six-axis force, end-effector pose, action commands, timestamps, and camera parameters. It also features hardware-level spatiotemporal synchronization, where all sensors are aligned to a unified physical coordinate system, explained RealMan Robotics.
The company highlighted five benefits that come with using multi-modal data:
- Ultra-low frame loss: Less than0.5% frame loss ensures continuous, reliable recording even at high speed.
- High-precision motion control: Millisecond-level joint data for smooth, accurate operations.
- Factory-calibrated for out-of-the-box use: No extra calibration is required.
- Generalization-oriented collection: Tasks can be repeated under diverse object, environment, and lighting conditions.
- Exoskeleton teleoperation: The dataset provides 1:1 human-to-robot motion mapping for high-fidelity demonstration.
Moving forward, the company plans to continue expanding the dataset, adding scenarios and modalities, and building a fully open, interconnected ecosystem that bridges research and industrial deployment.
Tell Us What You Think!