Motion data is commoditizing. Trust is not.
We make robot motion data signed, licensed, and traceable — who performed it, under what consent, with which genuine sensor channels. Built on open standards, starting with Japanese craftsmanship.
Vision-language-action foundation models increasingly learn motion directly from video. Open X-Embodiment publishes over one million real robot trajectories for free, and the unit cost of high-quality teleoperation data keeps falling. The motion file is a depreciating asset.
What does not depreciate: knowing who performed the motion, whether they consented, whether the force and tactile channels are genuinely measured, and whether the record is unaltered. As generation gets cheap, verification becomes the scarce layer. We are building that trust layer — not another file host.
Just as music went from something you attended to something licensed and played everywhere, physical skill is becoming something that is captured, signed, licensed, and installed on robots. A sushi master’s grip, a choreographer’s routine, a veteran fitter’s assembly sequence — knowledge that lived in a single body, replayed on machines worldwide, with the performer’s rights intact.
What that world requires is the equivalent of the music industry’s master-rights and publishing layer: whose performance is this, who holds the licence, and to whom does revenue flow each time it is trained on or executed. That layer does not yet exist for motion data. Formats (LeRobot) and models are already converging in the open — what is missing is the rights and trust infrastructure. That is the seat we are taking.
We apply the same discipline that runs our hiring-momentum dataset — primary-source capture, append-only storage, and full methodology disclosure — to motion data. We do not invent a new format: data ships in LeRobot-compatible form, and we own only the verification layer above it.
| Layer | Question It Answers | Mechanism |
|---|---|---|
| 🔏 Provenance | Is this record authentic? | Device-attested signatures at capture time; append-only lineage with no retroactive edits |
| 📜 Rights | Am I allowed to train on this? | Performer consent verified; per-skill licences with revenue share to the original performer — the music-publishing model, applied to body knowledge |
| 🖐 Quality | Does it carry what video cannot? | Declared and verified force/torque/tactile channels; de-duplication and explicit gap reporting |
The unit of trade is neither a video nor a raw log — it is a signed skill package. Each package contains:
Crafts disappearing with an ageing generation of artisans are captured before retirement — with consent and revenue share — and passed on to successors, schools, and robots from the same package.
Choreographers license routines that robots and avatars perform worldwide — legitimate distribution where every replay pays the rights holder, instead of unconsented scraping.
“This task, on this line, in this factory” exists nowhere on the internet. Sites capture and own their work, turning it into training assets for their own robots.
With a ground-truth motion, a learner’s movement can be diffed against it. Beyond robots that perform: robots that teach — the master’s motion licensed together with instructional annotations.
We start with motion data that has no alternative supply source anywhere in the world: the handwork of sushi artisans, traditional Japanese craft techniques, and precision manual assembly. The force and contact knowledge that never appears in internet video — captured with consent, signed, and licensed.
Supply scales in three tiers. (1) Individual performers — beyond artisans: professional athletes, dancers, and the motion actors behind games and film. Commercial mocap is a buyout business with zero royalties, so per-replay revenue share answers a long-standing grievance of the performers themselves. (2) Studio archives — Japan’s game and film industries hold some of the world’s deepest motion-capture archives; we work with rights holders to license them properly. (3) Commissioned site captures — recording site-specific tasks in factories and stores.
No bespoke lab is required. IMU suits (Rokoko, Xsens), optical stages (OptiTrack, Vicon), mocap gloves, and the force/tactile sensors that measure what video cannot (6-axis F/T, distributed tactile skin) — all off-the-shelf. Our proprietary layer is only the step that hashes and device-signs the recording at capture time. The full capture stack is available for purchase, rental, or lease through Smartmart, the Physical AI equipment store operated by our group.
| Phase | What Happens |
|---|---|
| Phase 0 — Now | Concept published; early-access intake. Design conversations with labs, model developers, and rights holders |
| Phase 1 — Pilot Captures | Consent-based pilot captures with artisans and performers. The signed-package spec is finalised on real data and provided to labs under limited terms |
| Phase 2 — Licensing API | Per-skill licence issuance, verification, and revenue split, automated. Buyers pull rights-cleared data directly from their training pipelines |
| Phase 3 — Marketplace | As capable robots reach scale, a two-sided market for listing, buying, and remixing skills — performers who teach while they sleep |
ASI operates a production alternative-data product in the Physical AI vertical (Hiring Momentum — fully disclosed methodology), practising primary-source collection, append-only storage, and data-lineage management daily. We also run a production pipeline for device-signed verifiable records — the foundation for motion-capture provenance signing.
Across the group, we run the full loop: capture hardware supply (Smartmart — mocap, 3D scanners, teleoperation kits, force sensors), signing and licensing of the motion data (this initiative), and the sale and rental of the robots that execute it. Hardware → capture → rights → replay on machines — owning every stage of the flywheel is what separates us from one-off data vendors.
Robotics labs, model developers, rights holders, and investors — we welcome conversations at the design stage. We typically reply within two business days.