Inconsistent sources and versions
Devices, software versions, and channels use different structures and topics.
Real-Robot Data
Move complex real-robot data into training faster
Real-robot data comes from physical systems, real tasks, and real execution pipelines. Modalities, topic structures, device versions, and failure cases make it difficult to use consistently.
KeenTruth profiles, governs, annotates, and validates customer-supplied real-robot data.
01
Data we handle
- Teleoperation data
- Leader–follower arm data
- Single- and dual-arm tasks
- Autonomous execution
- Robot demonstrations and policy replay
- Video, ROS bag, MCAP, HDF5, and custom structures
02
Common problems
Missing or interrupted records
Missing topics, corrupted files, and incomplete task segments reduce sample integrity.
Unclear task boundaries
Long sequences lack task hierarchy and action phases.
Invalid segments
Failures, ineffective actions, and device anomalies contaminate normal samples.
03
KeenTruth services
- File, modality, and topic profiling
- Completeness and anomaly checks
- Valid-data filtering and batch organization
- Semantic annotation and task decomposition
- Object boxes and hand keypoints
- Dynamic, temporal, and third-party output QA
04
Suitable training tasks
- Robot policy learning
- Task understanding and planning
- VLA model training
- Manipulation recognition
- Replay and failure analysis
- Data selection for model iteration
FAQ
Frequently asked questions
Can data from multiple devices or vendors be normalized?
Yes. We analyze source differences first, then define unified governance and output standards.
Do customers need to preprocess ROS bag or MCAP first?
No. Customers can provide raw data, a data dictionary, and task information.
Can you identify anomalies without annotation?
Yes. Real-robot governance and independent QA can be delivered as standalone services.
Contact
Move raw embodied data into training faster.
Tell us about your data, annotation requirements, and quality expectations.