Robotics · Imitation Learning Data
Sourcebae captures head-mount, egocentric, and exocentric demonstration data - time-synced, calibrated, and annotated - to train robot manipulation policies at scale.
Most manipulation models learn best from complementary views. We run all three modalities — independently or rig-synced on the same take — so you choose the coverage your architecture wants.
Lightweight head rigs let demonstrators move and work hands-free — ideal for mobile and bimanual tasks where a fixed camera can't follow.
The world as the demonstrator sees it: hand–object interaction, occlusion, and intent captured from the point of action.
Calibrated multi-camera rigs observe the scene from the outside, giving full-body context and a third-person frame robots can map to.
Egocentric frames keep the manipulated object and the hand in view through occlusion, so policies learn what the operator actually attends to and does.
Exocentric views expose the same task from arbitrary angles, helping policies transfer to a robot's own camera placement instead of overfitting one POV.
Time-synced ego + exo on a single take is the substrate for view-invariant representations, retargeting, and sim-to-real alignment.
Every take ships as aligned multi-modal data with the metadata your training pipeline expects — not raw footage you have to clean.
30–60 FPS, ego & exo
Per-frame, registered
Accel + gyro, 200 Hz
2D/3D keypoints
Fixation + attention
Segmented, timestamped
Intrinsics + extrinsic
Shared clock, μs-level
We map your tasks, environments, and target schema, then design the modality mix and capture protocol.
Head rigs and external cameras are set up, calibrated, and clock-synced to a shared timeline.
Trained operators perform demonstrations across the variations you specify — objects, lighting, locations.
Streams are aligned to a shared clock and reviewed for coverage, calibration drift, and label quality.
Hand pose, gaze, action segments, and contact events are labeled and reviewed against your guidelines.
Packaged in your format with calibration and metadata, shipped in batches with a documented spec.
We run capture as an operation: trained operators, repeatable rigs, and QA built in - so volume goes up without quality going down.
Trained operators
Hrs demo data / month
Modes | Head · Ego · Exo
Hour pilot turnaround
Robot imitation learning data consists of recorded human demonstrations that show how physical tasks are performed. These datasets can include synchronized video, depth, motion, gaze, hand pose, contact events, action segments and task metadata for training and evaluating robot manipulation policies.
©Sourcebae 2026 | All Rights Reserved
