Custom data production, built around your model.
Scenario development, recruiting, environment builds, capture, licensing, QC, and delivery — all in China, all to your spec.
Scope a pilotThe data your model needs next doesn't exist yet.
We don't sell a warehouse of stale data. We produce what your model is missing.
We build the supply chain that produces it
Requirements → paid pilot → scaled production. Priced per trajectory, per valid hour, per accepted task, or per environment.
We bridge the gap between your model's needs & China's data supply
One counterparty on your contract; behind it, a local network of studios, operators, recruiters, and engineers producing to a single schema and quality bar.
WE HANDLE
Scoping a pilot around your model’s weakest task
Static data, offline RL, or online RL: choosing the right deliverable
From 200 trajectories to production: how pilots scale into standing supply
RGB, depth, state, or semantic actions: nailing the schema first
Pricing models: per trajectory, per valid hour, per accepted task
Built for the labs training tomorrow's agents, world models, and robots.
What a paid pilot delivers.
50
tasks per paid pilot (20–50), designed around your model's weakest capability.
200
trajectories per pilot (50–200), delivered against an agreed schema.
100%
of delivered data ships with a documented license chain and quality report.
6+
synchronized streams per trajectory: video, inputs, state, task, result, failure reason.
4 weeks
typical time from scoping call to pilot delivery with samples and schema.
“Tell us the task your model keeps failing. We design the tasks, recruit the operators, build or license the environment, and deliver verified trajectories.”
For World Model Labs
GAME & SIM DATAThe data race is on.
Everyone is scaling compute. Trajectories are the bottleneck.
Explore Data ProductsWeb text is exhausted. The next capability gains come from data that records actions and their consequences — and the labs that lock in supply first hold an advantage that is hard to copy.
Three Capabilities.One Supply Chain.
Local sourcing & rights
We negotiate directly with Chinese studios and contributors — rights diligence, AI-training licenses, and recruiting handled locally, in Mandarin, at local cost.
Engineering integration
Engine hooks, capture rigs, and environment builds — we do the integration work so trajectories arrive synchronized and schema-complete.
Verification & QC
Task design, automatic result verification, and human QC — every batch is validated before it ships, and priced only when it passes acceptance.
One pipeline from scoping call to delivery
Step 1–2: Scope & pilot
We confirm what you need — static data or interactive environments, SL, offline or online RL, which signals, which actions — then run a paid pilot of 20–50 tasks with samples, a schema, and a quality & licensing report.
Step 3: Scale production
Once the pilot passes acceptance, production runs on standing infrastructure — priced per trajectory, valid hour, accepted task, or environment, plus development and licensing fees.
Built to Be Audited.
Training data is only as good as its provenance. Every delivery documents where the data came from, who authorized it, and how quality was verified — so your legal team signs off as fast as your research team does.
Licensing & Provenance
Every dataset ships with a documented license chain: who owns the game, who consented to capture, and what the AI-training grant covers.
Verified Quality
Automatic result verification plus human QC on every batch. You pay per accepted task, per valid hour, or per trajectory — not per attempt.
Flexible Deliverables
Static datasets, task generators, or executable RL environments — RGB, depth, state, raw or semantic actions, in whatever schema your stack expects.
