Results our clients ship with

The challenge

A humanoid robotics company building general-purpose industrial robots needed richly annotated perception data to teach their systems to grasp objects and navigate cluttered work cells safely. Their data spanned multiple modalities — RGB video, LiDAR point clouds, and depth — and factory environments are unforgiving: reflective metal surfaces, partial occlusion behind machinery, tightly packed parts, and humans moving through shared space where a mislabeled frame becomes a safety risk.

What we did

Vidyut Data stood up a multimodal annotation team trained on the client's object taxonomy and safety conventions. We delivered 3D bounding boxes and semantic segmentation on fused LiDAR + camera scenes, instance masks for graspable parts and tools, and skeletal keypoints for the humanoid's own limbs and nearby human workers — all temporally tracked across sequences and cross-checked between sensors. Ambiguous cases like reflective bins and occluded pallets were escalated, adjudicated, and written into a living labeling guide.