Training data for the tasks robots can't do yet
We capture what cameras miss — force, contact, recovery — and deliver synchronized, quality-checked episodes ready for your training pipeline.

FROM CAPTURE TO DEPLOYMENT
Every episode passes through four stages before delivery
CAPTURE
Multi-camera egocentric video, LiDAR depth, and tactile force sensing. Skilled workers performing real manipulation tasks. Synchronized to sub-10ms.
ANNOTATE
Automated pre-labeling — hand pose estimation, object segmentation, action classification — on every frame. Expert human review on the edge cases.
VALIDATE
QA gate. Modality sync within one frame, dropout and NaN detection on every channel, smoothness checks on hand pose and action vectors. Failures and recoveries included by design.
DELIVER
LeRobot v3 format. GR00T N1 compatible. Drop it into your training pipeline. Delivered via bucket or direct transfer.
4
Synchronized modalities per episode
LeRobot v3
Native format
MCAP/HDF5
Industry-standard formats
GR00T
N1 compatible
FAQ
LeRobot v3 (synchronized MP4 video + Parquet state/action files). Compatible with GR00T N1, Isaac Lab, and standard VLA training pipelines. Delivered via bucket or direct transfer.
Egocentric RGB video, LiDAR depth, tactile force sensing, 3D hand kinematics (MANO format), and language task descriptions. All synchronized to sub-10ms.
Every episode passes a QA gate: sync verified within one frame, NaN and dropout detection on every channel, velocity-spike checks on pose and action vectors. Anything that fails halts the run.
Precision assembly, tool use, textile manipulation, surface quality inspection, and more. Each domain captures contact-rich, bimanual tasks where vision-only systems fail.
Yes. Failures and recovery behaviors are captured alongside successes. Models need to learn what goes wrong and how to correct.
Contact us. Tell us what task domains you need and what's missing from your current training corpus. We'll scope it from there.