Subtask spans
Per-frame interval labels mapping every episode to a hierarchy of named subtasks (reach → grasp → lift → place).
Convert raw robot, simulation, and egocentric episodes into subtask spans, action primitives, contact events, success/failure labels, reward signals, and QA metadata. Run by trained robotics annotators inside the SignIQ Platform.
What we annotate
Per-frame interval labels mapping every episode to a hierarchy of named subtasks (reach → grasp → lift → place).
Atomic-action labels (open/close gripper, retract, pour, push) aligned to action chunks for ACT/Diffusion-style training.
Per-episode target object lists, bounding boxes when needed, and hand-object contact events for VLA reasoning supervision.
Onset timestamps for tactile array contacts and wrist F/T threshold crossings — synchronized to RGB and proprioception.
Outcome label per episode plus failure-mode tag with operator-rated severity (0→5 scale) for contrastive training and eval gating.
Per-step or sparse reward signals for offline RL, preference modeling, and reward-aware behavior cloning recipes.
Hand pose, finger articulation, gesture probabilities, and IMU streams aligned to first-person video for human-to-robot transfer.
Subtask-span alignment between human egocentric clips and matched robot teleop episodes with alignment-confidence scores.
12-check schema validation, sync drift, joint-limit, NaN/Inf, calibration provenance, and per-episode trainability scores.
Workflow
Episodes flow in from your captures or a SignIQ collection run.
Motion-energy + action-magnitude segmentation proposes subtask boundaries.
VLM hindsight captions, action-primitive classifiers, and contact detectors run as a first pass.
Trained robotics annotators correct boundaries, fix mislabels, and verify outcome tags.
Automated 12-check pipeline runs per episode; failures route back to review or are re-collected.
Delivered LeRobot, GR00T-LeRobot, HDF5, or MCAP — schema-validated for your training stack.
Platform connection
The marketing site explains what we annotate. Project setup, operator assignment, review queues, QA dashboards, and exports happen inside the protected workspace. Pilots typically scope to one task family and 100–500 episodes.