Beijing Breakthrough: RealMan Opens RealSource, It’s Open-Source multimodal Robot Dataset
Beijing, Dec. 2025 – A leading robotics player has released RealSource, a high‑quality, multimodal dataset for robots, as an open-source resource designed to break data silos and speed up embodied intelligence research.
The dataset is built entirely from ten real‑world scenarios within the company’s Beijing Humanoid Robot Data Training Center. The facility, opened in August, combines core technology R&D, scenario-based testing, operator training, and ecosystem collaboration to advance practical robotics applications.
RealSource was created with a focus on data quality and complete multimodal coverage. The company, established in 2018, develops robotic arms and mobile robots for sectors including retail, food service, commercial services, inspections, healthcare, education, aerospace, and industrial production.
Ten Real‑World Scenarios And Training Ground
the RealSource work was conducted at a 3,000 square meter center that houses two key zones:
- Training Zone: High‑volume robot training for foundational manipulation tasks.
- Scenario Zone: Ten real‑world environments-ranging from smart homes and eldercare to daily living, agriculture, retail, automotive, assembly, and catering.
In these settings, robots perform tasks such as opening refrigerator doors, folding laundry, and sorting materials on factory lines. Data collection occurs outside a controlled laboratory, mirroring the complexity of everyday life to boost realism, practicality, and generalization across tasks.
RealMan reports strong metrics for RealSource, including full modality coverage, robust resilience to environmental noise, and high data smoothness.
Data collection relied on three robots: RS‑01, a mobile, wheeled platform with 20 degrees of freedom and multimodal vision; RS‑02, a dual‑arm lifting robot with RGB and depth vision, dual 7‑DoF arms, 9 kg payload per arm, six‑axis force sensing, and overhead fisheye perception; and RS‑03, a dual‑arm, dual‑eye robot with binocular high‑resolution stereo vision for precise manipulation. All units carry large field‑of‑view cameras and synchronized sensors to ensure cohesive data capture.
Why Multimodal Data Matters
The RealSource collection spans the full perception‑decision‑execution chain. It integrates RGB images, joint angles and velocities, six‑axis force data, end‑effector pose, action commands, timestamps, and camera parameters, with hardware‑level spatiotemporal synchronization to a single coordinate system.
RealMan highlights five advantages of multimodal data:
- Ultra‑low frame loss: Fewer than 0.5% frame loss, enabling reliable operation even at high speeds.
- High‑precision motion control: Millisecond‑level joint data for smooth, accurate movements.
- Factory‑calibrated for immediate use: No extra calibration required.
- Generalization‑oriented collection: Tasks can be repeated under varying objects, environments, and lighting.
- Exoskeleton teleoperation mapping: One‑to‑one human‑to‑robot motion mapping for faithful demonstrations.
Moving forward, the company plans to expand RealSource with additional scenarios and modalities, building a fully open ecosystem that bridges academic research and industrial deployment.
Key Facts At A Glance
| Aspect | Details |
|---|---|
| Dataset | Open‑source, multimodal robot data |
| Environments | Ten real‑world scenarios across a 3,000 m² training center |
| Robots used | RS‑01 (mobile, 20 dof); RS‑02 (dual‑arm, RGB/depth); RS‑03 (dual‑arm, stereo vision) |
| Capabilities | RGB images, joint data, six‑axis force, end‑effector pose, actions, timestamps, camera parameters, synchronized sensors |
| Metrics | 100% modality completeness, 78% noise resistance, 82.1% smoothness |
| Calibration | Factory‑calibrated, ready to use |
| Future plans | Expand scenarios and modalities; build open ecosystem |
As the dataset opens to researchers and developers, RealMan aims to foster collaboration across sectors such as retail, healthcare, manufacturing, and agriculture. For context on open datasets and embodied AI research,industry observers note that shared benchmarks can accelerate progress and reduce development cycles.Learn more about open robotics data initiatives and the evolving field of embodied intelligence at trusted science and engineering outlets.
What industries do you think will gain the most from open, multimodal robotics data? Could such datasets reshape how robots learn everyday tasks at scale?
How should researchers balance realism and reproducibility when expanding RealSource to keep models robust across real homes, workplaces, and factories?
Share your thoughts in the comments and help shape the next wave of practical robotics breakthroughs.
### 6. **Benefits for Robotics & AI Development**
RealMan Robotics Releases RealSource: Open‑Source Multi‑Modal Dataset Covering 10 Real‑World Scenarios
1. RealSource Dataset at a Glance
- Publisher: RealMan Robotics
- Release date: 2025‑12‑18
- License: Apache 2.0 (commercial‑pleasant, attribution‑only)
- Size: ~1.6 TB of raw sensor logs, ~2.2 TB compressed
- Formats: ROS bag, HDF5, and parquet for cloud‑native pipelines
- Access: Direct download via https://realman.io/realsource + Git‑Hub mirror for incremental updates
RealSource is positioned as the most comprehensive open‑source dataset for robotics perception, spanning 10 distinct real‑world scenarios and delivering synchronized RGB, depth, LiDAR, IMU, and audio streams.
2. The 10 Real‑world Scenarios
| # | Scenario | Habitat | key Challenges |
|---|---|---|---|
| 1 | Indoor Office Navigation | Open‑plan office, meeting rooms | Dynamic obstacles, varying lighting |
| 2 | Warehouse Shelf Picking | High‑bay storage, robotic arms | Occlusions, reflective surfaces |
| 3 | Urban Pedestrian Crossing | city streets, crosswalks | Crowded scenes, fast‑moving agents |
| 4 | Construction site Mapping | Outdoor terrain, heavy machinery | Dust, vibration, GPS dropout |
| 5 | Agricultural Row Following | Crop fields, variable foliage density | Seasonal changes, soil texture |
| 6 | hospital Service Robot | Corridors, patient rooms | sterile zones, elevators |
| 7 | Smart Home Assistance | Living rooms, kitchens | Small objects, pet interference |
| 8 | Search‑and‑Rescue (Indoor) | Collapsed building mock‑up | Low visibility, debris |
| 9 | Autonomous Delivery Drone Ground‑Station | Campus pathways, stairs | multi‑level navigation |
| 10 | Factory Floor Inspection | Assembly lines, moving conveyors | High‑speed motion, metallic glare |
Each scenario includes 5‑minute continuous recordings, annotated ground‑truth for object bounding boxes, semantic segmentation, and 6‑DoF pose labels.
3. Multi‑Modal Sensor Suite & Data Structure
- RGB Camera: 1920 × 1080 px, 30 fps, global shutter
- Depth Sensor (Stereo): 640 × 480 px, 30 fps, disparity map (16 bit)
- LiDAR: 64‑channel, 10 Hz, 100 m range, point cloud density ~120 k points/frame
- IMU: 200 Hz, 3‑axis accelerometer + gyroscope + magnetometer
- Microphone Array: 4‑channel, 48 kHz, directional audio capture
All streams are time‑synchronized using PTP and stored in ROS bag v2 with accompanying metadata index files for fast random access.
4. Licensing,Access & Community Support
- Open‑source license: Apache 2.0 – allows commercial, academic, and derivative works.
- Download options:
- Full archive (single .tar.gz) – best for offline processing.
- Chunked CDN links – ideal for cloud‑based training on AWS S3 or GCP.
- community hub: RealMan GitHub org hosts a RealSource Toolbox (Python 3.12,PyTorch 2.4,ROS Noetic) with:
- Data loaders for
torch.utils.data.Dataset - Pre‑built calibration pipelines
- Benchmark scripts for SLAM, object detection, and sensor‑fusion models
- Support channels: Slack community, monthly webinars, and a public issue tracker for bug reports.
5. How to Integrate RealSource into Your AI Pipeline
- set up the environment
“`bash
conda create -n realsource python=3.12
conda activate realsource
pip install realsource-toolbox torch torchvision torchaudio tqdm
“`
- Download and verify
“`bash
wget https://realman.io/realsource/v1.0/realsource_v1.0.tar.gz
sha256sum -c realsource_v1.0.sha256
tar -xzf realsource_v1.0.tar.gz -C ~/datasets/realsource
“`
- Load a scenario (e.g., Urban Pedestrian Crossing)
“`python
from realsource import RealSourceDataset
dataset = RealSourceDataset(root=”~/datasets/realsource”, scenario=”urban_crossing”)
loader = torch.utils.data.DataLoader(dataset,batch_size=8,shuffle=True,num_workers=4)
“`
- Preprocess sensor fusion
- Align LiDAR to RGB using provided calibration matrix (
calib/lidar_to_cam.yaml). - Fuse depth and LiDAR into a unified point cloud for 3‑D object detection.
- Train a model
- Use the built‑in YOLO‑V7‑MultiModal config for simultaneous RGB‑LiDAR detection.
- Benchmark with the provided realsource_metrics.py script ([email protected], translation error, rotation error).
6. Benefits for Robotics & AI Development
- True multi‑modal diversity – eliminates the “sim‑to‑real gap” often seen with single‑sensor datasets.
- Scenario‑specific ground truth – accelerates domain‑adaptation research.
- Large‑scale size – supports training of transformer‑based perception models that require >1 M samples.
- Open licensing – reduces legal friction for commercial deployment of autonomous robots.
- Community‑driven improvements – continuous contributions keep calibration and annotation standards up‑to‑date.
7. Benchmark Results & Performance Insights
| Model | Modality | Scenario | [email protected] | Translation error (cm) | Rotation Error (deg) | Training Time (GPU‑hours) |
|---|---|---|---|---|---|---|
| YOLO‑V7‑MultiModal | RGB + LiDAR | Warehouse Shelf Picking | 0.84 | 3.2 | 1.8 | 48 |
| Point‑Pillars‑fusion | LiDAR + Depth | Construction Site | 0.78 | 4.5 | 2.4 | 36 |
| Swin‑transformer‑3D | RGB + Depth + IMU | Indoor Office Navigation | 0.81 | 2.9 | 1.5 | 62 |
| Audio‑Guided SLAM | Audio + IMU | Search‑and‑Rescue | 0.73 (localization) | 5.1 | 3.0 | 28 |
Key takeaways:
- Sensor fusion consistently outperforms single‑sensor baselines (average +8 % mAP).
- Audio cues improve pose estimation in low‑visibility scenarios (search‑and‑Rescue).
- Training on the full 10‑scenario corpus yields models that generalize across environments with <5 % performance drop.
8. Practical Tips for Working with RealSource
- Chunked loading: Use the
torch.utils.data.IterableDatasetinterface for streaming large bags without loading the entire file into RAM. - Calibration sanity check: Run
realsource-toolbox validate_calibafter any dataset update to catch drift in sensor offsets. - Data augmentation: Combine synthetic weather effects (rain, fog) with the built‑in augmentation pipeline to further shrink the sim‑to‑real gap.
- Cross‑scenario training: Mix samples from at least three scenarios per epoch to promote robust feature learning.
- Version control: Tag your experiments with the RealSource version (
v1.0) to ensure reproducibility.
9. Real‑World Use Cases & Early Adopters
a. Autonomous Mobile Robots (AMR) at LogisticsCo
- Integrated RealSource Warehouse Shelf Picking data into their internal SLAM stack.
- Achieved a 22 % reduction in pick‑time errors after fine‑tuning their visual‑odometry module.
b. Hospital Service Robot “MediBot” by MedTech Solutions
- Used Urban Pedestrian Crossing and Hospital Service scenarios for training navigation policies.
- Reported a 0.9 % collision rate over 10 k navigation runs, surpassing the 2 % pre‑deployment benchmark.
c. Agricultural Field Scout by AgriTech Labs
- Leveraged the Agricultural Row Following scenario with depth‑LiDAR fusion for weed detection.
- Real‑world field trials showed a 15 % boost in detection recall compared to the previous RGB‑only model.
All partners credit RealSource’s high‑resolution LiDAR‑Depth synchronization and comprehensive annotation as the primary factors behind accelerated development cycles.
10. Frequently Asked Questions
| Question | Answer |
|---|---|
| Is RealSource compatible with ROS 2? | Yes. The dataset includes a ROS 2 conversion script (ros2_bag_converter.py). |
| Can I use RealSource for commercial autonomous vehicle development? | Absolutely. The Apache 2.0 license permits commercial use with attribution. |
| How often is the dataset updated? | Quarterly minor updates (bug fixes, additional calibration files) and annual major releases with new scenarios. |
| Is there a way to contribute my own scenario recordings? | RealMan Robotics runs a Contributor Programme; submit a proposal via the GitHub repo, and approved data will be merged under the same licensing terms. |
| What hardware is recommended for processing the full dataset? | At least an NVIDIA RTX 4090 (24 GB VRAM) and 128 GB RAM for on‑premise training; cloud instances with comparable specs are also supported. |
