Japanese worksite data for physical AI development.
We record egocentric video and the tacit knowledge behind skilled work at sites in Japan where people are actually working, with a first focus on industry and infrastructure. Capture follows your model specification, at the volume you need.
Start with a conversation about your data specification.
Director: Toshiyuki Yamamoto, founder of Chatwork
01 / PROBLEM
You have the model. The field data is missing.
01
Real-world data sets the ceiling
Simulation and public datasets take a policy only so far on real tasks. Data recorded in real environments is what decides how a model performs in the field.
02
Public datasets do not differentiate
When every team can train on the same corpus, it is hard to build an edge on it. Differentiation is moving toward data that only your team can access.
03
Opening up worksites is not R&D work
Securing consent to film, clearing rights, building trust with the people doing the work. The stage before collection is a different job from model development.
02 / WHY EGOCENTRIC
Why egocentric video, and why tacit knowledge.
01
Manipulation data does not exist on the web
Language models learned from web text, and image models learned from web images. Operational data from real work is not on the web: it has to be recorded one session at a time in the physical world. The quality and diversity of those demonstrations is what ultimately decides how imitation learning and VLA models behave on site.
02
The egocentric view is close to what a robot has to reproduce
A skilled worker's head-mounted view records the relationship between gaze, both hands, tool and workpiece as it happens. Occlusion, gaze shifts and fine hand detail, all lost to a fixed camera, stay in the recording. And because no robot hardware is involved, capture can go where the work actually happens instead of into a lab.
03
Skill does not show up on video alone
Why the worker stopped there. What they looked at before calling it good or bad. Judgment is hard to learn from footage alone, so we interview the worker and attach their intent as timestamped annotation, letting behavior and decision criteria be learned together.
Public datasets such as Open X-Embodiment and DROID are collected largely in research settings. Egocentric data recorded at manufacturing and service sites in Japan is available only to a limited extent in public form.
03 / DATA
What we deliver
EGOCENTRIC VIDEO
Egocentric video
Head-mounted capture of a skilled worker: gaze, hands and tool operation. The world as the operator sees it, which is the view a robot has to reproduce.
TACIT KNOWLEDGE
Verbalized tacit knowledge
Why the worker paused here. What they looked at before deciding. We interview the worker, turn intent and decision criteria into annotation, and deliver it linked to the video.
CUSTOM COLLECTION
Collection to your specification
Camera setup, resolution, frame rate, synchronization, metadata. We design and collect against a written specification aligned with your model and training pipeline.
What the data specification fixes
Before collection starts, we fix the following in a written data specification. We can propose a setup based on how the data will be used: pre-training, fine-tuning or evaluation.
- Viewpoint and rig
- Head-mounted camera placement, count and mounting method; optional hand-level or fixed auxiliary views
- Video
- Resolution, frame rate, field of view, and how stabilization is handled
- Sync and metadata
- Time synchronization across viewpoints; metadata fields such as environment, material, tooling and working conditions
- Annotation
- Task segmentation granularity (process, task, motion), format for tacit-knowledge annotation (intent and decision criteria), pass/fail labels
- Delivery
- File format, folder structure and delivery method; sample delivery, your review, then full collection
- Rights
- Exclusive or non-exclusive license, scope of use (training, evaluation, redistribution, derivatives), scope of consent from the site and the worker
We set licensing — exclusive or non-exclusive — and the scope of use per contract, to match your intended use.
04 / FIELD ACCESS
We source sites in Japan to match your specification.
We are not limited to a single industry. You tell us the work you want captured and the conditions it has to meet; we look for a matching site in Japan, clear the rights to film and to use the footage, and then record. That sourcing and clearing is the part we take on. We are glad to talk before the target task is settled.
FOCUS
Industry and infrastructure
This is where we are focusing first. Hand skill decides the outcome, and passing that skill on is an open problem across much of this work. We are starting from in-factory manufacturing work and designing capture around it.
Example tasks we expect to capture
Construction and building services
Work where the procedure shifts with each individual object, such as installation and connection.
Retail and hospitality
Work with parallel task flows, such as a restaurant kitchen and floor.
Household and domestic work
Work where conditions differ in every home and the sequence is planned on the spot, such as domestic help.
Agriculture
Work where season and weather change the conditions, and the next action is chosen by reading how the crop is growing.
These are areas we expect to cover. They are not sites we own or operate. For each engagement we source a site in Japan that matches the specification at that time.
A site that has agreed to filming
A factory producing steel water pipes and formed fittings has agreed to filming of manual welding and gouging. We are preparing a sample collection there.
Not a mock research setup, but sites where people are actually working.
05 / PROCESS
From first conversation to delivery
- 01
Discovery
We go through the target tasks, how the data will be used, and the specification you have in mind. It is fine if the specification is not settled yet.
- 02
Data specification
We fix camera setup, annotation format and granularity, and delivery format in a written document.
- 03
Sample collection
We collect a small volume so you can check the content against the specification.
- 04
POC collection
We collect and deliver at volume against the agreed specification.
- 05
Ongoing supply
We keep supplying data while widening the range of sites and tasks covered.
Pricing is quoted individually once the data specification is fixed.
06 / COMPLIANCE
Rights clearance is part of data quality.
Data with unclear rights cannot be used for training. We settle this in writing with the site before collection begins.
Written agreement with the site
Filming and data-use rights are agreed in writing with the operating company before collection starts.
Consent from the worker
We obtain consent from each worker on the use of their likeness and on the scope of data use.
Care for trade secrets
We manage what appears on camera and can mask areas that need it.
NDA
We can sign an NDA from the specification discussion onward.
07 / COMPANY
Company
- Company name
- Chuo-Sogo Inc.
- Founded
- January 29, 2026
- CEO
- Shion Seki
- Director
- Toshiyuki Yamamoto (founder of Chatwork)
- Location
- 1-5-18-3 Shinoharakita, Kohoku-ku, Yokohama, Kanagawa 222-0026, Japan
- Business
- Collection and supply of training data for physical AI / website production
08 / FAQ
Frequently asked questions
Camera position and count, resolution, frame rate, the target work, annotation format and granularity, and delivery format, agreed against a written specification. It is fine to come to us before the specification is settled.
It depends on what is collected, the volume, and the license model (exclusive or non-exclusive), so we quote individually once the data specification is fixed.
We deliver only data for which rights have been cleared in writing with both the operating company and the worker. The scope of use is set out in the contract.
We are not limited to a single industry. Our first focus is manufacturing work in industry and infrastructure, and we also expect to cover construction and building services, retail and hospitality, household work, and agriculture. Our role is to find a site in Japan that matches the specification for each engagement, clear the rights to film and to use the data, and record. If the work you need is not listed, tell us what you want captured and under what conditions.
Public datasets such as Open X-Embodiment and DROID are collected largely in research settings, and the same data is available to every other team. We record at sites in Japan where people are actually working, to your specification, and attach tacit-knowledge annotation drawn from interviews with the worker.
You can start with a small sample collection. The sequence is specification alignment, sample collection, your evaluation, then POC collection, so you can check the content as you go.
Task segmentation (process, task, motion), tacit-knowledge annotation (intent and decision criteria drawn from worker interviews), pass/fail labels, and other formats defined in the specification. We can also deliver in the format your existing pipeline uses.
Camera setup, resolution, frame rate, synchronization method and delivery format can be specified on your side. We can also collect using hardware and formats you provide.
09 / CONTACT
Contact
Get in touch about a data specification, a sample collection, or anything else you want to discuss. We reply within two business days.