Optics
what the hands seeEgocentric video with 6DoF head pose, hand-skeletal tracking, and depth. Episode-level annotation with task segmentation. X hours across X countries.
1618 captures how people actually move through the world: what they see, hear, do, and feel, recorded as one synchronized signal. Robots trained on a single sense are apprentices. Robots trained on all of them are polymaths.
A mechanical knight, driven by cables and pulleys, able to sit, stand, and move its arms. He built it by studying human anatomy first and machinery second. He understood that to make a machine move like a person, you must first measure how a person moves.
Five hundred years later, this is still the problem. The internet taught language models everything humans have written. It taught robots nothing about how humans work. There is no archive of a hand adjusting its grip on a wet glass, no record of the force a wrist applies when a drawer sticks.
We are building that archive. All senses at once, measured from the body itself.
Optics, acoustics, anatomy, and mechanics shared the same notebooks five centuries ago because their author understood them as one subject. So do we. The 1618 Standard binds the four streams into one synchronized session, time-locked at the source.
Egocentric video with 6DoF head pose, hand-skeletal tracking, and depth. Episode-level annotation with task segmentation. X hours across X countries.
In-environment audio: verbal instruction, tool contact, material response, ambient context. Multilingual, recorded where work actually happens.
Whole-body pose, 7-DoF arm kinematics, joint angles, and IMU streams captured on the same rig as the video. Motion recorded as measurement, not estimated from pixels afterward.
Force, resistance, and compliance from instrumented grippers during teleoperated manipulation. The data that tells a policy how hard is hard enough.
Three capture tiers feed the notebooks: head-mounted egocentric recording, handheld instrumented grippers, and teleoperation. On request, the same task is captured across all three tiers: paired data for teams comparing training recipes.
A dataset assembled from separate collections of video, audio, and kinematics is a collage. The timing between what a person sees and how their grip responds is lost, and that timing is the training signal.
The 1618 Standard requires every modality on a shared clock with hardware-level timestamps. What you receive is not four datasets zipped together. It is one recording of a human performing a task, complete.
Some robots learn by copying demonstrations. Others learn a model of the world from continuous experience, and plan with it. The first wants clean episodes: one task, one reset, one instruction. The second wants long, uncut footage, because every cut destroys the thread of time it learns from.
1618 records one way and delivers both. Contributors capture long continuous takes; task boundaries and scene resets are marked, never cut. You order episodes, or you order the whole river. Same footage, your packaging.
For world-model teams there is a third product: grounding packs. Small volumes of manipulation data where every action is time-locked to what the camera saw, held to the Standard's strictest synchronization clause.
Leonardo's Vitruvian Man made the human body the unit of measurement. 1618 does the same for robot training data.
Every dataset we release is scored on the Vitruvian Benchmark, our open evaluation framework for egocentric manipulation data. The methodology, the reference samples, and our scores publish together with our first release. Score anyone's data with it, including ours.
| Metric | What it measures | Score |
|---|---|---|
| Action density | Manipulation time as a share of total footage | X |
| Scene diversity | Distinct environments, objects, lighting regimes | X |
| Annotation fidelity | Label accuracy against double-blind human review | X |
| Kinematic completeness | Pose and joint coverage across the session | X |
| Temporal continuity | Cut-free certification: unbroken time per take | X |
| Temporal sync error | Cross-stream timestamp deviation, milliseconds | X |
Each recording carries contributor ID, capture environment, hardware serial, consent record, and timestamp.
Contributors are verified, compensated, and credited. Faces and bystanders are handled by a redaction pipeline that runs on our own hardware.
Terms are explicit: commercial training use, licensee-owned weights. Exclusive commissions are enforced in our delivery ledger, not just in the contract. Deliveries can be locked to a collection region; the pipeline refuses violations.
1.618 is the golden ratio, the proportion Leonardo found throughout the human body, the oldest standard of human measurement in existence. We took the name because it describes the job. Sample packs for evaluation: X hours, all four modalities, benchmark scores included. Custom collections scoped within X days.
Request samples