Real factory environments
Captured on live production lines, not in synthetic or staged setups.
Third Origin · Human workflow data
Access the most exhaustive source of ready-labelled, end-to-end multimodal workflow data from diverse real-world manufacturing environments.
Thirty task specimens · ego + exo · RGB–D · UMI bimanual · instrumented gloves · annotation demo · library map
01 · Platform
Third Origin provides structured, training-ready workflow data from real manufacturing environments. Models trained on it see long-horizon tasks the way workers actually do them, decision points included.
Captured on live production lines, not in synthetic or staged setups.
Multi-step tasks with their transitions and decision points, not isolated pick-and-place.
Labeled, formatted, ready to load. Raw video alone does not train anything.
Policies train on the mess they will actually meet, which shrinks the sim-to-real gap and holds up over long horizons.
02 · Task specimens
Thirty tasks from working factories across a dozen sectors. Each note says what makes the task hard for a robot to learn.
01 glass works
Corrugated sleeves over fragile stock; grip force gets no second try.
02 appliance line
Small device, moving belt, both hands busy at once.
03 assembly bench
Identical parts, tiny registration marks, zero tolerance for drift.
04 industrial qc
Count, check, log: perception and bookkeeping in one loop.
05 garment line
Snips against soft knit; one wrong cut shows forever.
06 knitting line
Yarn tension by feel, machine timing by ear.
07 packaging line
Flat blank to box in seconds, creases from muscle memory.
08 ceramics works
Unfired clay marks at a touch; pressure is the whole skill.
09 electronics bench
Checkerboard, module, micro-adjust: sub-millimetre alignment, on camera.
10 pleating bench
A straightedge, a knife, and folds that must repeat exactly.
11 moulding bench
Each cavity takes one part, seated flush or not at all.
12 assembly line
Long flexible part, narrow channel, continuous press-fit.
13 lampshade bench
Accordion folds that collapse if the order is wrong.
14 wax works
Rubber bands around soft blocks: deformation on every grip.
15 garment line
The same fold every time; the pile is the QC.
16 leather works
Blade angle by feel; one pass decides the edge.
17 garment line
A drawstring through a channel, blind for most of the run.
18 electronics bench
Cluttered bench, live device, probe and screen in parallel.
19 footwear line
Layered soft parts aligned by hand before the press.
20 stainless works
Gloved work under a running head; force where the camera can't see.
21 upholstery bench
Pry, stretch, seat: a stiff panel into a tight housing.
22 lockworks
A bowl of near-identical brass parts, picked by touch.
23 appliance assembly
A harness routed through housings, connectors seated blind.
24 cleanroom lab
Gloved pipette work into microtubes; contamination is failure.
25 metal shop
Rivets and sheet stock: high force, exact placement.
26 glass works
Wet, clear, slippery, and invisible to depth sensors.
27 packing floor
Tension a strap around a soft bundle without crushing it.
28 footwear line
Machine-paced stitching where two curved parts meet.
29 print shop
Ruler-true cuts on stacks that shift as they square.
30 garment line
A press machine, a placement mark, and no undo.
03 · What ships
The data foundation for physical AI: real tasks and real decisions, captured at scale.
5,000+ real manufacturing tasks, with the natural variation and edge cases that make generalization possible.
Data includes atomic actions and high-level human commentary, so one dataset teaches low-level control and task-level reasoning.
Synchronized ego + exo video, tactile, and contact signals for learning grounded physical interaction and fine-grained manipulation.
Ships as structured datasets with temporal alignment and segmentation, ready to drop into a training pipeline.
04 · Reach
We work with factories that let us put cameras directly on the line: garments, cosmetics, packaging, electronics, mining, industrial tooling. Each sector brings its own tools, its own materials, and its own ways of going wrong.
05 · Diversity
Explore 900 captioned windows from the same 30 hand-selected tasks shown in the carousel and specimen catalog. Search for a manipulation, inspect the cluster, then open the published clip that produced it.
t-SNE projection of four-second window captions. Axes carry no units; nearby points describe similar work.
06 · Annotation · 190 steps · 21-point hands
A two-minute ironing line with every step labelled and per-frame hand keypoints. Step boundaries come from the hand kinematics, not the model; the curve on the timeline is the motion signal itself.
Annotation timeline
07 · Why us
Define the spec with us, see early captures, and adjust before the full collection runs.
Robotics and world-model teams shape what we capture, so the data answers training questions, not filming ones.
Captured directly from live manufacturing environments: real tools, materials, constraints, and edge cases.
Task structures, action hierarchies, and annotations defined with partners, aligned to model training objectives.
Every modality passes multi-stage QA before delivery. If a stream is unusable, you never see it.
Every dataset is rights-cleared and auditable, and we adapt handling to your compliance requirements.
08 · Products
Off-the-shelf labeled datasets
Custom data collection
Thirty tasks across glass, ceramics, garments, electronics, footwear, metal, and packaging work, each with a note on why it is hard for a robot.
Watch the clips →Filter assembly, ironing, arm assembly, parcel labeling, and pin insertion: 789 labelled steps with per-frame hand keypoints, playing in sync.
Open the viewer →900 captioned windows from the 30 curated task specimens, searchable by what the hands are doing and traceable to the published clips.
Search it above →Workflow data that teaches models whole jobs, not isolated actions.