Syngraph captures expert human demonstrations of contact‑rich manipulation tasks, synchronizing video, depth, and tactile sensor data. Each episode is automatically pre‑labeled, reviewed by experts, and passes a seven‑check QA gate before being delivered in the LeRobot v3 format, compatible with GR00T N1, Isaac Lab, and standard VLA pipelines for immediate use in robot learning workflows.
Funding
Funding not disclosed
Founders
Product
Problem
Robotic manipulation research lacks large-scale, high-fidelity datasets that capture both visual and tactile information of skilled human hand work, making it difficult to train models for precision assembly, cable routing, and textile manipulation.
Solution
Syngraph records expert human demonstrations of contact-rich tasks, synchronizing video, depth, and tactile sensor data. Each episode undergoes automated pre‑labeling (hand pose, object segmentation, action classification) followed by expert review of edge cases. A rigorous seven‑check QA gate validates modality synchronization, detects dropouts, NaNs, and ensures smoothness of pose and action vectors. Validated episodes are packaged in the LeRobot v3 format, compatible with GR00T N1, Isaac Lab, and standard VLA pipelines, and delivered via cloud bucket or direct transfer for immediate integration into training workflows.
Target Audience
Primary customers are foundation model teams building large‑scale pretraining datasets and deployment teams requiring contact‑rich, high‑precision manipulation data for robot learning pipelines.
Features
- Automated pre‑labeling of hand pose, object segmentation, and action classification on every frame
- Expert human review for edge‑case validation ensuring high annotation quality
- Seven‑check QA gate covering frame‑level sync, NaN/dropout detection, and velocity‑spike checks on pose and action vectors
- Synchronized multimodal capture of video, depth, and tactile force data
- Delivery in LeRobot v3 (MP4 video + Parquet state/action files) with native support for GR00T N1 and industry‑standard formats (MCAP/HDF5)
- Flexible data transfer via cloud bucket or direct pipeline integration