This dataset has no comparable source. Catalog access: non-exclusive by default. Exclusivity available on specific trades categories.
Construction-Native Training Data for Manipulation Policies
Egocentric 4K video + wrist IMU from skilled tradespeople. Every clip is annotated, expert-validated, anonymized, and legally documented. Ships natively in LeRobot v3.0.
No construction-domain egocentric data exists at scale
Annotation Density
Per-frame
Every 0.5s: action, tool, grasp, force, intent
What Ships With Every Dataset
Per-Frame Annotations
Action labels, tool detection, material identification — COCO-compatible bounding boxes at 2Hz across every clip.
MANO Hand Pose
21-keypoint hand tracking per hand, synchronized with wrist IMU at 100Hz. Grasp type classification included.
Optical Flow + Depth Maps
Dense motion fields and estimated depth for spatial reasoning. Derived from stereo and structure-from-motion.
Expert-Validated Intent
Every annotation reviewed by the skilled tradesperson who performed the task. Not just labeled — confirmed by the person who did it.
Error Recovery Sequences
Deliberately captured failure modes: cross-threading, mis-cuts, alignment errors — and the expert recovery that follows.
AI-Guess Delta
Every annotation includes the disagreement between AI inference and worker correction. Alignment training data for physical AI hallucinations.
The AI-Guess Delta
Every annotation includes the disagreement between AI inference and tradesperson correction. When our model said "tightening at 40 Nm" and the pipefitter corrected it to "65 Nm, cross-threaded, backing out" — that delta is in the data. This is alignment training data for physical AI hallucinations: the exact gap between what a model predicts and what a skilled human actually observes.
Pipeline-Ready Format
Ships natively in LeRobot v3.0 format — drops directly into your training pipeline. COCO-compatible bounding boxes, MANO keypoints, IMU CSV with hardware-synced timestamps. No conversion scripts, no format wrangling, no surprises.