Data Annotation Services

Raw teleoperation recordings need structured labels before they become training-ready datasets. SVRC provides frame-level annotation, task segmentation, quality scoring, and natural language labeling — so your policy training pipeline gets clean, well-structured inputs.

Annotation Capabilities

Frame-Level Labeling

Object bounding boxes, semantic segmentation masks, keypoint annotations, and contact state labels on every camera frame. Supports multi-view consistency checks across synchronized camera streams.

Task Segmentation

Temporal boundary annotation that splits long demonstrations into meaningful sub-tasks and skill primitives. Essential for hierarchical imitation learning, skill chaining, and language-conditioned policy training.

Quality Scoring

Per-episode quality scores based on trajectory smoothness, grasp success, task completion, and operator consistency. Identify and filter low-quality demonstrations before they contaminate your training set.

Additional Label Types

  • Success / failure tagging — binary outcome labels per episode and per sub-task, with failure mode classification (missed grasp, collision, timeout, incorrect placement)
  • Natural language descriptions — free-text and templated task descriptions for language-conditioned policies (SayCan, RT-2, OpenVLA style)
  • Reward shaping signals — dense reward annotations for reinforcement learning from demonstrations (RLfD) pipelines
  • Object state tracking — per-frame object pose, articulation state, and contact graph annotations

Process

You send us raw episodes (HDF5, RLDS, LeRobot, or video + state logs). We define the annotation schema together, train annotators on your specific task taxonomy, run annotation with double-blind QA, and deliver labeled data in your target format. Typical turnaround is 3–10 business days depending on volume and label complexity.

All annotation work is managed through the Fearless Data Platform, which provides real-time progress tracking, annotation review interfaces, and direct export to your training pipeline.

Need your robot data annotated?

Send us a sample episode and your annotation requirements. We will return a scoping document and timeline within 48 hours.