Robotics
Annotation
Physical-world training data is scarcer and costlier to collect than anything digital. We label it across synchronised sensor streams — pose, grasp, kinematics, and the safety events that decide whether a system ships.
A robot episode is not a video.
Video annotation gives you boxes on frames. A robot episode is several sensor streams that must first be time-aligned, then labeled together — because the grasp only makes sense alongside the force reading that accompanied it.
That means the ingest layer alone is heavier than most annotation pipelines: ROS bags, LiDAR sweeps, depth frames, IMU, and force-torque traces arriving at different rates in different formats. Getting that synchronisation right is a precondition for any label being meaningful.
Safety labeling compounds it further. Collision events and human-robot interaction violations are rare, high-consequence, and easy to miss — exactly the class of label where a published agreement floor matters most.
What We Annotate
Eight labeling disciplines spanning perception, manipulation, imitation learning, and physical safety.
Multi-sensor ingestion
ROS bag parsing with time-synchronisation across LiDAR, depth camera, IMU, and force-torque streams arriving at different sample rates.
3D scene & pose annotation
Object pose estimation, grasp-point labeling, semantic segmentation for navigation, and scene-graph annotation of spatial relationships between objects.
Trajectory & kinematics
Joint-angle sequences and end-effector paths over time — the physical-world counterpart to agent trajectory labeling, on the same underlying schema.
Human demonstration annotation
Teleoperation and kinesthetic teaching capture, labeled for imitation learning — a category growing fast as robotics teams shift to learning-from-demonstration.
Sim-to-real validation
Validates synthetic data from Isaac Sim, MuJoCo, or Gazebo against real-world behavior, so simulator-trained policies survive contact with physics.
Safety & collision labeling
Structured labeling of unsafe behavior, collision events, and human-robot interaction safety-zone violations — mapped to ISO 10218 and ISO/TS 15066.
Temporal action segmentation
Segments continuous operation logs into discrete labeled tasks and sub-tasks using change-point detection plus human review.
Physical-world ontology
A standardized object, material, and affordance taxonomy — what can be grasped, pushed, opened — so labels stay model-consistent across every project you run with us.
The same trajectory model powers our agentic AI work.
A robot arm choosing a grasp and a software agent choosing a tool are the same shape of problem — state, action, outcome, over time. One schema, one timeline viewer, one export pipeline serves both. Teams building embodied agents get a single annotation standard across the digital and physical halves of the system.
See agentic AI annotationBuilding robots? Let's talk about the data.
We're running the robotics annotation pillar with a small group of design partners — teams who want to shape the tooling as it's built, and get priority capacity in return.