/ COLLECT · PRODUCT

Robot data, captured clean and labeled.

Demonstration, teleoperation, and autonomous sensor data — time-aligned across every modality, quality-checked, auto-annotated, and exported in the format your training stack expects.

1M+
trajectories collected
16
sensor modalities
99.9%
pipeline uptime
<24h
demo → dataset

/ WHAT WE CAPTURE

Every modality, on one clock.

No more stitching streams together after the fact. Telemanual records every signal time-synchronized to the same trajectory.

Multi-camera video

Synchronized RGB and depth from wrist, scene, and overhead views.

Proprioception

Joint positions, velocities, torques, and end-effector poses.

Force & tactile

6-axis force/torque and tactile arrays for contact-rich tasks.

Audio & events

Ambient and contact audio, plus discrete event markers.

Language labels

Natural-language instructions and per-segment annotations.

Custom streams

Any extra sensor — bring a topic, we time-align it.

/ COLLECTION MODES

Three ways to generate data, one schema out.

Teleop demonstrations

Expert operators drive the robot; every session is a clean demo.

Autonomous rollouts

Capture policy executions in the loop to grow your dataset.

Scripted protocols

Repeatable scripted tasks for benchmarking and coverage.

/ QUALITY PIPELINE

Labeled and QA'd before it ever hits your bucket.

Auto-annotation runs inline — segmentation, success/failure tags, language instructions, and contact events. Human review catches the rest.

Automatic success / failure classification
Segmentation and contact-event detection
Language instruction tagging
Human-in-the-loop spot review & QA
ingest_pipeline · running
14:22:01sync 4 camera streams + joints
14:22:03auto-label: pick → place (success)
14:22:04detect contact events ×3
14:22:06convert → LeRobot + MCAP
14:22:07sync → s3://your-bucket
2,148 / 2,560 trajectories formatted → s3://your-bucket

Get your first labeled trajectories this week.

Start a pilot